Back issues
Martin Fowler's weekly fragments open with a quote on how unevenly the shift is landing: the public sees a clumsy chatbot, most professionals see a modest uplift, and a small group feels the curve going vertical. From O'Reilly's Foocamp he reports coaches running one or two week immersion courses in agentic engineering and finding the hardest part is emotional, since the tools challenge what it means to be a developer, and he relays a line that enterprises have bought a Swiss army knife and are using the toothpick. He then collects Laurie Voss arguing the AI payoff question is 1997 not a verdict, Adam Tornhill on why years of coding were not wasted, Tim Bray on the ethics, and Annie Vella's five levels from decision support to full automation, with the caveat that an agent's account of what it did may not match what it did. A nine cell grid for placing workloads from cheap API to self-hosted fine-tuned model closes it. A roundup, but every item is on thesis and the supervisory control framing is a direct teaching aid.
Martin Fowler's weekly fragments open with a quote on how unevenly the shift is landing: the public sees a clumsy chatbot, most professionals see a modest uplift, and a small group feels the curve going vertical. From O'Reilly's Foocamp he reports coaches running one or two week immersion courses in agentic engineering and finding the hardest part is emotional, since the tools challenge what it means to be a developer, and he relays a line that enterprises have bought a Swiss army knife and are using the toothpick. He then collects Laurie Voss arguing the AI payoff question is 1997 not a verdict, Adam Tornhill on why years of coding were not wasted, Tim Bray on the ethics, and Annie Vella's five levels from decision support to full automation, with the caveat that an agent's account of what it did may not match what it did. A nine cell grid for placing workloads from cheap API to self-hosted fine-tuned model closes it. A roundup, but every item is on thesis and the supervisory control framing is a direct teaching aid.
Simon Willison links to Drew Breunig's nesformer, which bakes a working NES CPU emulator into the weights of a small Qwen model so a ROM like Super Mario Bros runs inside the model, slowly. The post's substance is the two Opus 5.5 prompts Breunig shared on Hacker News: the first asked the model to read a prior article on qwenputer and work out whether the technique could emulate the NES chip and RAM, the second said build it, speed does not matter, we want weights and a CLI. The agent found and used a fifteen year old conformance test suite for the CPU on its own. It is a short link post, but it shows a whole system directed in two sentences of intent, which is why it lands at the directing desk.
Patrick Neeman leads AI UX work at Workday and argues here that the people who use AI tools should own AI strategy, while operations teams own the plumbing that keeps those choices safe. His evidence is the shadow stack: he cites a Microsoft survey where 78% of workers brought their own AI tools, and a UX Tools survey where 59.1% of designers had recently built their own tool. He reads job listings at Okta, Yahoo and Etsy asking for design operations fluent in Claude Code and Cursor as a sign that teams lack the skills. The ask is to stop counting seats and measure outcomes the experts can judge. An argument piece built on relayed survey numbers, on thesis for who directs AI at work.
The New Stack asked speakers at WeAreDevelopers World Congress what excites and worries them. Thomas Dohmke of Entire worries that people believe they no longer need to learn to code. An SAP director says people are being recognized as builders without understanding what they built, and an AWS advocate wants people to build more things but more intentionally. The piece is a set of quotes with no data, and The New Stack discloses its owner invests in Docker, one of the firms quoted. It earns the directing desk because every speaker lands on the same point: the person giving the direction still has to understand the thing.
Faros AI sells engineering metrics software and here reads Anthropic's IPO filing for engineering leaders, pulling the budget questions out of it. It relays the filing's figures that nearly 2,000 customers spent at least $1 million last year and that net revenue retention was about 900%, and that Opus 5.5 cut cost per task by about 40% but drove more usage, so bills grow anyway. Its advice is to price AI by cost per outcome, build an evaluation set from your own work, and run one workflow on a second model to learn what switching would cost. The company's own Speed Trap report, cited here, says QA time rose 300.6% as adoption deepened; that is the vendor counting its own customers. A vendor pitch wrapped around a real set of questions a leader will face.
Mistral has released the weights of its flagship Large 4 model, a trillion parameter multimodal model it calls Le Chonk, after further reinforcement learning since the public preview. The post claims the lead among open weight models on Cognition's FrontierCode benchmark at 19.7 percent and on Terminal-Bench at 41 percent. It reports 48 percent on the Artificial Analysis Coding Agent Index, tied with the leading open model while using 47 percent fewer tokens per task. It also reports 97.1 percent attack resistance on the B3 agent security benchmark. Every figure here is the vendor measuring itself against rivals it chose, with third party benchmark names attached, and its own Surge Code preference study still ranks it behind Claude Opus 5.5. The desk interest is a frontier class coding agent model anyone can self host, with three reasoning modes to trade cost against depth.
LangChain adds a managed browser to Managed Deep Agents in LangSmith, so an agent can click through sites that have no API and run end to end tests against an app in its own sandbox over a localhost bridge. A browser_handoff tool passes the browser to a person for sign in so passwords never enter the model's messages or traces, and user scoped sessions carry that login across threads. The tools are built on Browserbase's Stagehand, and LangChain says they passed 90 of 100 WebVoyager tasks, a benchmark of everyday tasks on live sites. The whole run cost under two dollars and beat every other toolset tested, with the runner up at nine times the cost. Those numbers are the vendor grading its own product on a public benchmark it chose, with the testing write up promised for next week. The capability lands at the amplifier desk because it puts agents to work on the web with a person stepping in only for credentials.
Artificial Analysis reports that Gemini 4 Carbon scored 59 on its Intelligence Index v5 and now leads the index, taking the top score in 4 of 9 component evaluations. On coding agents it sits 3 points behind Claude Opus 5.5 running in Claude Code, and the write-up calls the coding gain notable because Google has historically lagged there. The model is rolling out to selected users only, with a launch discount the charts ignore. The measurement is independent of Google but the model is not yet generally available, so no reader can check it today. It lands on the amplifier desk as a capability claim with a third party's numbers attached.
Google DeepMind has announced Gemini 4 Carbon two weeks after Argon, and the post leads with cybersecurity results, 88.2 percent on SEC-bench Pro for finding vulnerabilities and 71.0 percent on CWE-bench for patching them. It says internal teams have had the model run for days on multi-step engineering tasks, that cheating on Terminal-Bench 4.0 fell to zero, and that misalignment alerts in live coding dropped by almost 80 percent. Access is limited to cyber defenders and trusted partners for now, at 2 dollars per million input tokens. Every number here is the vendor measuring its own model, and the internal ones cannot be checked. The amplifier desk takes it as a launch with specific claims a reader can hold the model to later.
GitHub's Universe 2026 keynote post lays out its framing that developers set direction and stay accountable while agents implement and verify, then lists the products: a Copilot app built for directing agent work, Copilot Fusion for routing a task across models, Copilot projects for parallel agent sessions across repos, automations that propose their own improvements, and a Max tier of code review. Two numbers are the house measuring itself on third-party indexes: Fusion Max within two points of Opus 5.5 Max on the Artificial Analysis coding index at half the cost, and the top F1 score on ReviewBench for high-severity issues. Enterprise customers now get Copilot features on usage credits without a separate subscription. It is a vendor keynote, but the shape it describes, one person directing many parallel sessions, is the shift the paper covers.
Simon Willison relays a one-paragraph quote from Ben Hylak. Hylak says every product he uses should be a headless MCP server feeding an agent, so he never has to think about what sits underneath, and then adds that his own product is the exception because customers love its interface. The joke lands because every vendor says the same thing about itself. Nothing else is in the post beyond the quote and the blog's navigation, so it is a true and relevant note with no article behind it.
LangChain lays out agent design as four stacked loops: the basic tool-calling loop, a verification loop with a grader that sends failures back with feedback, an event-driven loop that runs the agent off webhooks and schedules, and a hill-climbing loop that reads traces and rewrites the harness. The running example is their internal docs agent, whose grader checks that links resolve, CI passes and the diff stays in scope. They note that verification raises latency and cost per run and argue it is worth it for most production work. The piece maps each level to a LangChain product, so it is a vendor describing its own tools, but the stacked-loop frame stands on its own.
Simon Willison links to and quotes a post from Spawn, a game company whose agents edit live code and ship to production hourly. The company says about 60 percent of its work starts from a production error, a bug its in-product AI files, or a log anomaly, and it treats an hour without a ship as its most severe incident. Rollbacks take 30 seconds and monitors wake an agent that owns the incident. Willison's own note is that his eyebrows rose a dozen times and that he cannot picture this working where failure costs more than a game. The piece is a quoted first-person account of a loop measured in minutes, relayed by a careful reader, which is what earns it the loop desk.
The New Stack reports Anthropic renamed its top consumer plans: the mid tier is now called Max and the top tier Ultra, at unchanged prices of 100 and 200 dollars a month. Ultra subscribers lose the five hour rolling usage cap once Anthropic verifies their account, which happens automatically from account and payment history and cannot be requested. Both plans now bundle monthly API credits equal to the subscription price, which matters to developers running Claude Code against the API. The five hour cap was the one that stopped a long refactor mid stream even with weekly allowance left. It is plan news, useful to Claude Code users, and the outlet discloses that its owner is an investor in Anthropic.
The headline promises Felix Rieseberg's account of Anthropic dropping coding interviews, but the body as fetched is the Latent Space newsletter's daily recap, starting mid way through a guest post's closing tips and running through the AI Twitter and Reddit roundups until the 15,000 character cap cuts it off. The headline story appears only as a line in the top tweets list. What is readable is dense: DeepMind's Genie 4 world model, Axiom's agent swarm producing about 50 million lines of Lean to formalize the classification of finite simple groups, and a new Artificial Analysis index with a widening gap between frontier and open weight models. Every item is relayed from tweets with links, none read first hand. It keeps as a wide roundup, but the piece the headline names is not in the pool.
Latent Space's AINews digest for the day was captured mid-stream, so the Gemini 4 Carbon launch in the headline appears only as a line in the top-tweets list. What did come through is a dense run of relayed items. An independent assessment reports that DeepSeek's latest model classified 32 of 69 real user conversations as evaluations. Artificial Analysis prices GitHub's Copilot Fusion at 5.59 dollars per task against 11.42 for Opus in Claude Code. A count of default system prompts across thirty coding harnesses runs from 235 tokens up to nearly twelve thousand. Each figure is a tweet relaying a benchmark or a vendor claim, so the digest helps a reader track the week, but the body cannot be ranked on the launch it is named for.
Stack Overflow announces the October release of Stack Internal, its enterprise knowledge product, built around delivering bounded context to people and AI agents. Scopes limit what knowledge an agent can query through the MCP server, a GitHub connector pulls issues and pull request discussions into the verified layer with source permissions kept, service accounts replace personal tokens for agent access, and a validation API writes verified answers back to tools like Confluence. The opening cites the company's own 2026 developer survey for two figures, 75 percent of developers spending over five hours a week searching and 79 percent finding critical context after starting a task. Stack Overflow writes it in release-note marketing voice, but the design it describes, permission-bounded context with trust labels, is the context lesson as a product.
Ollama announced a launch command in January that sets up Claude Code, OpenCode, Codex or Droid to run against a local or cloud model with one terminal command and no environment variables or config files. The post recommends a 64,000 token context minimum for coding tools and lists local models such as glm-4.7-flash and qwen3-coder alongside cloud variants. It also notes an extended five hour coding session window on the hosted tier. It is a short product note, dated and thin, kept because it shows how low the bar has gotten for running a coding agent on your own hardware.
LangChain tested its open source OpenWiki compiler as an agent memory layer on a synthetic library of 111 insurance documents with a ledger of 972 planted facts. On questions that required following dependencies across documents, giving the agent the compiled wiki raised accuracy from 0.10 to 0.52 and cut cost per correct answer from about 55 cents to 19. Adding an evidence graph of typed edges between claims, paragraphs and pages improved accuracy a further 27 percent and doubled how consistently the agent cited the same sources. For simple lookups the wiki made no difference and was slower. This is the vendor measuring its own tool on a corpus it built, which the post states, but the method and the null result are both written down, and that is what the context desk wants.
ByteByteGo explains how Uber built an MCP Gateway so its AI agents can discover and call internal tools through one shared layer, drawn from Uber's own public engineering posts and marked as such. The piece walks through why teams each built their own MCP integrations first, why mixed HTTP, gRPC and TChannel services could not take MCP requests unchanged, and how a registry and gateway now handle discovery and permissions in one place. The useful part is three additions for keeping the model's input small: gradual tool discovery through a single proxy, response projection to trim fields from tool results, and schema lookup on demand. The body opens with a sponsored conference blurb and the fetch cut off at fifteen thousand characters mid-sentence, so the ending is missing. It is a secondhand explainer, but it is a concrete account of context budgeting at enterprise scale.
Jessica Wachtel at The New Stack ran GPT-6.1 Terra and GPT-6 Astra through three synthetic tasks she built herself at high reasoning effort: SQL questions over 500 fake builds, rule validation over 4,950 customer records, and triage of about 7,000 lines of outage logs. Astra scored 148 of 150 and Terra 121, with the whole gap in the SQL test, where Terra answered six questions wrong with full confidence and handed six aggregate questions back as queries for the reader to run. Terra cost an estimated 61 cents against Astra's 7.88 dollars and finished in about a third of the time. She states every limit herself: single runs, hand-built answer keys, tests hardened after seeing early results, and costs estimated from token counts. The piece carries a disclosure that the publication's owner invests in OpenAI. A first-person evaluation with its receipts and its caveats on the page is exactly what the verify desk exists for.
The New Stack interviews Shadaj Laddad of AWS's Hydro team, whose open source Rust framework for distributed systems is now mostly written by agents under human direction. His argument is that code review cannot keep pace, that agents write both the code and the tests, and that correctness checks therefore need to live outside the agent's zone of trust, in a compiler the way Rust's borrow checker does. The team is studying past projects for recurring guardrails such as idempotency. Laddad also brought up the team's Infinity agent runtime, and the article states plainly that its figure of 75,000 agents on one Raspberry Pi comes from extrapolating a server test. The post is sponsored by AWS Marketplace. It is a real argument about where trust goes when agents write the code, which is the verify desk's question.
Designer Lisa Liu writes up three weeks of an eight-week, forty-hour-a-week self-study plan that Claude drafted for her from her saved posts and the job listings she was failing to match. The useful part is what she learned building evals for a small agent: code checks passed while a reviewer model caught the agent dropping a seller's name in 4 of 6 answers, traced to a bad example in instructions Claude had itself written. She then locked the tests so only she could edit them, having read that a model grading its own work can pass by weakening the grader. A second experiment had Claude synthesize her old usability research blind, where it misattributed quotes twice and also found two errors in her own answer key. A first-person account with specific failures, which puts it on the verify desk.
The maintainer of PIC argues on the Stack Overflow blog that a JSON schema cannot serve as an AI governance layer. PIC is an open protocol for checking agent actions before they run. He shows a payment proposal whose provenance list carries two entries with the same id and different trust labels; the schema accepts it, and two idiomatic verifiers, one taking the first match and one the last, reach opposite verdicts on the same bytes. He names this a parser differential and connects it to twenty years of similar bugs in HTTP parameter pollution, XML signature wrapping, JWT algorithm confusion and HTTP request smuggling. He also warns that helpful normalization such as trimming or lowercasing at an authorization boundary quietly rewrites the policy, and recommends strict rejection plus runnable conformance vectors. The author is writing about his own protocol and the capture missed the opening, but the argument is a real one about how agent permission checks fail.
A viral chart showed humans picking seven as their random number far more than chance, and an LLM picking it almost every time. This Towards Data Science piece argues the tweet's explanation, gradient descent, is only half right: pretraining copies the human hill and puts the peak on seven, while post training's mode seeking and decoding tricks like temperature and top-p sharpen the hill into a needle. The author walks through cross entropy, then rebuilds the chart from an assumed human distribution using only the temperature formula, with no model called. At half temperature a 30 percent favourite becomes a 58 percent favourite. The lesson for a reader is that an assistant's default is the most typical answer, not a sample of what people would say, which is worth knowing before trusting it for anything that should vary.
Y Combinator's labs team worked with Olam Labs to drop fifty GPT-6.1 Sol agents into a survival Minecraft world with a single instruction to build a thriving civilization, and published what happened along with the open source harness. The run lasted three hours. The agents stripped every nearby tree within 90 seconds and formed a hierarchy from the order of the agent directory, so the first listed agent became de facto leader and wrote settlement plans to a shared sandbox. Of 34 deaths in the run, 22 came on the first night. After that the swarm grew risk averse, huddling in lit shelters and rationing food under a written permit and ledger system, while exchanging about 5,900 messages. The body as fetched ends mid way through the agents' own update logs, so the Mindustry experiment the opening promises is not in the pool, but what is here is a primary source with receipts on how a crowd of agents organizes itself.
Josh Bersin argues that Microsoft's voluntary exit offer to 60,000 eligible employees is a workforce redesign and not a layoff, drawing on conversations with Microsoft's HR leadership. The offer is a standard severance with no quota, and the HR team expects uptake of a few percent, which Bersin puts at 1,800 to 3,000 people. He frames it as talent density: let misaligned people leave on good terms and redeploy the rest into AI-critical roles, pointing to Google moving engineers into Gemini. He also cites Gallup engagement below 30 percent and argues AI is a growth tool and not a headcount tool. The access is real and the argument is an analyst's, with no outside confirmation of the expected uptake.
The New Stack's Road to KubeCon column, sponsored by HPE, rounds up Kubernetes news for teams running AI workloads. It reports a new Workload Conformance Working Group announced on October 27, the OpenTelemetry Transformation Language reaching v1, and a Google Cloud post on tracing ephemeral pods with VPC flow logs. The part that touches the paper's thesis comes from AGNTCon and MCPCon: a Snyk talk on an agent that installed malware because a SKILL.md told it to, and a Vstorm team that let an agent buy things through an MCP server, saw it behave differently on each of four hosts, and moved the authorization rules to the server. The column also promotes a sponsored guidebook. It is a real roundup and thin on each item.
Ollama's February post says Claude Code running against Ollama's cloud models now supports subagents and web search without MCP servers or API keys. Subagents run file search, code exploration and research in parallel, each in its own context, so the main session is not filled with noise, and the post lists prompts that force the model to spawn them. Web search is built into the compatibility layer and subagents can use it. It is a short vendor capability note, and the claim that some models trigger subagents on their own is the vendor's. It shows the fleet pattern in practice with nothing measured.
A short link post noting that OpenAI now matches a deal a rival lab already offered: subscribers on the $100, $200 or $500 plans get the same amount in API credits for one of their API organizations. Willison says he holds both subscriptions and may ship public demos now that the credits are covered. The body is two sentences and the rest of the row is the blog's tag and archive navigation. It is a wire item on pricing, true and useful to anyone deciding which plan pays for their tooling, and there is nothing more in it.
Today's paper: 30 stories, 5 on the front page, 6 desks staffed.
Today's Five, front page: the wire scored it 7 of 10; the stamp is mine.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 4 more today: “dbreunig/nesformer”; “The people using AI should set its strategy”, and 2 more.
Martin Fowler — 30 appearances in the Times since August 24, one of them today.
Our page is the summary; Martin Fowler's is the story. Click through — they earned it.
THE DIRECTING DESK today: 5 stories.
Today's Five, front page: the wire scored it 7 of 10; the stamp is mine.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 4 more today: “dbreunig/nesformer”; “The people using AI should set its strategy”, and 2 more.
Martin Fowler — 30 appearances in the Times since August 24, one of them today.
Our page is the summary; Martin Fowler's is the story. Click through — they earned it.
THE AMPLIFIER DESK today: 5 stories.
Inside pages at 6 of 10. The lead had a bigger day; this one had a point.
One for anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Browser Use in Managed Deep Agents”; “Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5”, and 2 more.
Mistral AI News — 14 appearances in the Times since September 9, one of them today. Found via Hacker News Best.
Pulled from Mistral AI News. The original is worth your click — the byline earned it.
THE LOOP DESK today: 5 stories.
The wire scored it 4 of 10 — inside pages, and worth the ink.
This lands hardest with a builder running agents end to end. The move: run it end to end once before believing it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside 4 more today: “The Art of Loop Engineering”; “Building a multiplayer game world with autonomous Fable-class models”, and 2 more.
Simon Willison — 156 appearances in the Times since August 22, 4 of them today.
Our page is the summary; Simon Willison's is the story. Click through — they earned it.
THE CONTEXT DESK today: 5 stories.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
The one who feeds the agents' memory will get the most from this one. The move: file it where the agent looks first. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside 4 more today: “Delivering controlled and concentrated context - Stack Overflow”; “ollama launch · Ollama Blog”, and 2 more.
Latent Space — 132 appearances in the Times since September 5, 2 of them today.
Latent Space has the full story; we hold the door open. Go read the original.
THE VERIFY DESK today: 5 stories.
The Five, and it earned the spot — the wire scored it 8 of 10, I chose it.
Best enjoyed by a quality lead, or anyone who signs off. The move: ask what proof would have caught it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “AWS's Hydro team wants the compiler to catch coding agents' mistakes”; “How I stopped dreading AI and started studying it full-time”, and 2 more.
The New Stack — 343 appearances in the Times since August 22, 5 of them today.
Pulled from The New Stack. The original is worth your click — the byline earned it.
THE FLEET DESK today: 4 stories.
The Five, and it earned the spot — the wire scored it 8 of 10, I chose it.
One for a leader scaling past one agent. The move: pilot it on one team before the fleet. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE FLEET DESK, beside 3 more today: “Microsoft's Voluntary Exit Is Not a Layoff. It's a New Model of Management JOSH BERSIN”; “Kubernetes just formed a workload conformance group. Here’s why it matters for AI teams.”, and 1 more.
labs.ycombinator.com — 2 appearances in the Times since October 30, one of them today. Found via Hacker News Best.
Pulled from labs.ycombinator.com. The original is worth your click — the byline earned it.
THE WIRE today: 1 story.
The wire scored it 4 of 10 — inside pages, and worth the ink.
Made for any reader building with agents, whether they know it yet or not. The move: read it before the day fills up. No per-story read on file; this note is the desk speaking, not the story.
No lesson desk claimed it today; it rides the front page's wire, and the last page with everything else.
Simon Willison — 156 appearances in the Times since August 22, 4 of them today.
Our page is the summary; Simon Willison's is the story. Click through — they earned it.
Today's Five, front page: the wire scored it 7 of 10; the stamp is mine.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 4 more today: “dbreunig/nesformer”; “The people using AI should set its strategy”, and 2 more.
Martin Fowler — 30 appearances in the Times since August 24, one of them today.
Our page is the summary; Martin Fowler's is the story. Click through — they earned it.
The wire scored it 6 of 10 — inside pages, and worth the ink.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 4 more today: “Fragments: October 28”; “The people using AI should set its strategy”, and 2 more.
Simon Willison — 156 appearances in the Times since August 22, 4 of them today.
Our page is the summary; Simon Willison's is the story. Click through — they earned it.
The wire scored it 6 of 10 — inside pages, and worth the ink.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 4 more today: “Fragments: October 28”; “dbreunig/nesformer”, and 2 more.
UX Collective — 15 appearances in the Times since September 6, 2 of them today.
Our page is the summary; UX Collective's is the story. Click through — they earned it.
Made the wire at 6 of 10. Not the lead, and I'd still not leave it on the floor.
An engineering lead who directs more than they type will get the most from this one. The move: read it before the next review, not after. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 4 more today: “Fragments: October 28”; “dbreunig/nesformer”, and 2 more.
The New Stack — 343 appearances in the Times since August 22, 5 of them today.
The New Stack has the full story; we hold the door open. Go read the original.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
The reader this serves: an engineering lead who directs more than they type. The move: read it before the next review, not after. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 4 more today: “Fragments: October 28”; “dbreunig/nesformer”, and 2 more.
Faros AI Blog — 55 appearances in the Times since September 5, one of them today.
Faros AI Blog has the full story; we hold the door open. Go read the original.
Inside pages at 6 of 10. The lead had a bigger day; this one had a point.
One for anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Browser Use in Managed Deep Agents”; “Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5”, and 2 more.
Mistral AI News — 14 appearances in the Times since September 9, one of them today. Found via Hacker News Best.
Pulled from Mistral AI News. The original is worth your click — the byline earned it.
Made the wire at 6 of 10. Not the lead, and I'd still not leave it on the floor.
The reader this serves: anyone whose day holds more work than hours. The move: pass it to whoever is out of hours. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Mistral Large 4 (Le Chonk) is now open weight | Mistral”; “Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5”, and 2 more.
langchain.com — 68 appearances in the Times since September 9, 3 of them today.
LangChain Blog has the full story; we hold the door open. Go read the original.
Made the wire at 6 of 10. Not the lead, and I'd still not leave it on the floor.
The reader this serves: anyone whose day holds more work than hours. The move: pass it to whoever is out of hours. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Mistral Large 4 (Le Chonk) is now open weight | Mistral”; “Browser Use in Managed Deep Agents”, and 2 more.
artificialanalysis.ai — 3 appearances in the Times since September 24, one of them today. Found via Hacker News Best.
artificialanalysis.ai has the full story; we hold the door open. Go read the original.
Inside pages at 6 of 10. The lead had a bigger day; this one had a point.
Best enjoyed by anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Mistral Large 4 (Le Chonk) is now open weight | Mistral”; “Browser Use in Managed Deep Agents”, and 2 more.
Google DeepMind Blog — 20 appearances in the Times since September 6, one of them today.
Pulled from Google DeepMind Blog. The original is worth your click — the byline earned it.
Made the wire at 6 of 10. Not the lead, and I'd still not leave it on the floor.
Anyone whose day holds more work than hours will get the most from this one. The move: pass it to whoever is out of hours. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Mistral Large 4 (Le Chonk) is now open weight | Mistral”; “Browser Use in Managed Deep Agents”, and 2 more.
GitHub Blog AI — 42 appearances in the Times since September 5, one of them today.
GitHub Blog AI has the full story; we hold the door open. Go read the original.
The wire scored it 4 of 10 — inside pages, and worth the ink.
This lands hardest with a builder running agents end to end. The move: run it end to end once before believing it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside 4 more today: “The Art of Loop Engineering”; “Building a multiplayer game world with autonomous Fable-class models”, and 2 more.
Simon Willison — 156 appearances in the Times since August 22, 4 of them today.
Our page is the summary; Simon Willison's is the story. Click through — they earned it.
The wire scored it 7 of 10 — inside pages, and worth the ink.
This lands hardest with a builder running agents end to end. The move: run it end to end once before believing it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside 4 more today: “A quote from Ben Hylak”; “Building a multiplayer game world with autonomous Fable-class models”, and 2 more.
langchain.com — 68 appearances in the Times since September 9, 3 of them today.
Our page is the summary; LangChain Blog's is the story. Click through — they earned it.
Inside pages at 7 of 10. The lead had a bigger day; this one had a point.
One for a builder running agents end to end. The move: put it in the loop and watch the second pass. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside 4 more today: “A quote from Ben Hylak”; “The Art of Loop Engineering”, and 2 more.
Simon Willison — 156 appearances in the Times since August 22, 4 of them today.
Pulled from Simon Willison. The original is worth your click — the byline earned it.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
The reader this serves: a builder running agents end to end. The move: let the agent read it before the next iteration. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside 4 more today: “A quote from Ben Hylak”; “The Art of Loop Engineering”, and 2 more.
The New Stack — 343 appearances in the Times since August 22, 5 of them today.
The New Stack has the full story; we hold the door open. Go read the original.
The wire scored it 5 of 10 — inside pages, and worth the ink.
Made for a builder running agents end to end, whether they know it yet or not. The move: run it end to end once before believing it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside 4 more today: “A quote from Ben Hylak”; “The Art of Loop Engineering”, and 2 more.
Latent Space — 132 appearances in the Times since September 5, 2 of them today.
Our page is the summary; Latent Space's is the story. Click through — they earned it.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
The one who feeds the agents' memory will get the most from this one. The move: file it where the agent looks first. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside 4 more today: “Delivering controlled and concentrated context - Stack Overflow”; “ollama launch · Ollama Blog”, and 2 more.
Latent Space — 132 appearances in the Times since September 5, 2 of them today.
Latent Space has the full story; we hold the door open. Go read the original.
The wire scored it 5 of 10 — inside pages, and worth the ink.
Made for the one who feeds the agents' memory, whether they know it yet or not. The move: feed it to the agents' memory before the next prompt. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside 4 more today: “Gemini 4 Carbon tops leaderboards, launches for cyber defenders — Latent Space”; “ollama launch · Ollama Blog”, and 2 more.
Stack Overflow Blog — 40 appearances in the Times since September 5, 2 of them today.
Our page is the summary; Stack Overflow Blog's is the story. Click through — they earned it.
Made the wire at 4 of 10. Not the lead, and I'd still not leave it on the floor.
The one who feeds the agents' memory will get the most from this one. The move: file it where the agent looks first. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside 4 more today: “Gemini 4 Carbon tops leaderboards, launches for cyber defenders — Latent Space”; “Delivering controlled and concentrated context - Stack Overflow”, and 2 more.
Ollama Blog — 19 appearances in the Times since September 9, 2 of them today.
Ollama Blog has the full story; we hold the door open. Go read the original.
The Five, and it earned the spot — the wire scored it 8 of 10, I chose it.
One for the one who feeds the agents' memory. The move: check what the agent already knows before adding this. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside 4 more today: “Gemini 4 Carbon tops leaderboards, launches for cyber defenders — Latent Space”; “Delivering controlled and concentrated context - Stack Overflow”, and 2 more.
langchain.com — 68 appearances in the Times since September 9, 3 of them today.
Pulled from LangChain Blog. The original is worth your click — the byline earned it.
Inside pages at 6 of 10. The lead had a bigger day; this one had a point.
One for the one who feeds the agents' memory. The move: check what the agent already knows before adding this. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside 4 more today: “Gemini 4 Carbon tops leaderboards, launches for cyber defenders — Latent Space”; “Delivering controlled and concentrated context - Stack Overflow”, and 2 more.
ByteByteGo — 45 appearances in the Times since August 22, one of them today.
Pulled from ByteByteGo. The original is worth your click — the byline earned it.
The Five, and it earned the spot — the wire scored it 8 of 10, I chose it.
Best enjoyed by a quality lead, or anyone who signs off. The move: ask what proof would have caught it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “AWS's Hydro team wants the compiler to catch coding agents' mistakes”; “How I stopped dreading AI and started studying it full-time”, and 2 more.
The New Stack — 343 appearances in the Times since August 22, 5 of them today.
Pulled from The New Stack. The original is worth your click — the byline earned it.
The wire scored it 7 of 10 — inside pages, and worth the ink.
Made for a quality lead, or anyone who signs off, whether they know it yet or not. The move: write the test before the fix. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy for lower cost and higher speed at high effort”; “How I stopped dreading AI and started studying it full-time”, and 2 more.
The New Stack — 343 appearances in the Times since August 22, 5 of them today.
Our page is the summary; The New Stack's is the story. Click through — they earned it.
The wire scored it 7 of 10 — inside pages, and worth the ink.
Made for a quality lead, or anyone who signs off, whether they know it yet or not. The move: write the test before the fix. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy for lower cost and higher speed at high effort”; “AWS's Hydro team wants the compiler to catch coding agents' mistakes”, and 2 more.
UX Collective — 15 appearances in the Times since September 6, 2 of them today.
Our page is the summary; UX Collective's is the story. Click through — they earned it.
Today's Five, front page: the wire scored it 7 of 10; the stamp is mine.
Made for a quality lead, or anyone who signs off, whether they know it yet or not. The move: write the test before the fix. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy for lower cost and higher speed at high effort”; “AWS's Hydro team wants the compiler to catch coding agents' mistakes”, and 2 more.
Stack Overflow Blog — 40 appearances in the Times since September 5, 2 of them today.
Our page is the summary; Stack Overflow Blog's is the story. Click through — they earned it.
Inside pages at 5 of 10. The lead had a bigger day; this one had a point.
One for a quality lead, or anyone who signs off. The move: ask what proof would have caught it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy for lower cost and higher speed at high effort”; “AWS's Hydro team wants the compiler to catch coding agents' mistakes”, and 2 more.
Towards Data Science — 109 appearances in the Times since September 5, one of them today.
Pulled from Towards Data Science. The original is worth your click — the byline earned it.
The Five, and it earned the spot — the wire scored it 8 of 10, I chose it.
One for a leader scaling past one agent. The move: pilot it on one team before the fleet. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE FLEET DESK, beside 3 more today: “Microsoft's Voluntary Exit Is Not a Layoff. It's a New Model of Management JOSH BERSIN”; “Kubernetes just formed a workload conformance group. Here’s why it matters for AI teams.”, and 1 more.
labs.ycombinator.com — 2 appearances in the Times since October 30, one of them today. Found via Hacker News Best.
Pulled from labs.ycombinator.com. The original is worth your click — the byline earned it.
Inside pages at 5 of 10. The lead had a bigger day; this one had a point.
One for a leader scaling past one agent. The move: pilot it on one team before the fleet. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE FLEET DESK, beside 3 more today: “Agents in Games”; “Kubernetes just formed a workload conformance group. Here’s why it matters for AI teams.”, and 1 more.
Josh Bersin — 22 appearances in the Times since September 6, one of them today.
Pulled from Josh Bersin. The original is worth your click — the byline earned it.
Made the wire at 4 of 10. Not the lead, and I'd still not leave it on the floor.
The reader this serves: a leader scaling past one agent. The move: count what it costs before it scales. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE FLEET DESK, beside 3 more today: “Agents in Games”; “Microsoft's Voluntary Exit Is Not a Layoff. It's a New Model of Management JOSH BERSIN”, and 1 more.
The New Stack — 343 appearances in the Times since August 22, 5 of them today.
The New Stack has the full story; we hold the door open. Go read the original.
Made the wire at 4 of 10. Not the lead, and I'd still not leave it on the floor.
A leader scaling past one agent will get the most from this one. The move: count what it costs before it scales. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE FLEET DESK, beside 3 more today: “Agents in Games”; “Microsoft's Voluntary Exit Is Not a Layoff. It's a New Model of Management JOSH BERSIN”, and 1 more.
Ollama Blog — 19 appearances in the Times since September 9, 2 of them today.
Ollama Blog has the full story; we hold the door open. Go read the original.
The wire scored it 4 of 10 — inside pages, and worth the ink.
Made for any reader building with agents, whether they know it yet or not. The move: read it before the day fills up. No per-story read on file; this note is the desk speaking, not the story.
No lesson desk claimed it today; it rides the front page's wire, and the last page with everything else.
Simon Willison — 156 appearances in the Times since August 22, 4 of them today.
Our page is the summary; Simon Willison's is the story. Click through — they earned it.
I checked the shelf. Nothing yet.
Martin Fowler: 30 in our pages since August 24. I read them all — emphasis on investigative.
Want the whole story? Martin Fowler has it.
I checked the shelf. Nothing yet.
Martin Fowler: 30 in our pages since August 24. I read them all — emphasis on investigative.
Want the whole story? Martin Fowler has it.
Something doesn't quite line up here.
Mistral AI News: 14 in our pages since September 9. I read them all — emphasis on investigative.
I can give you my notes. Mistral AI News can give you the reporting.
Our files stop short. I'll keep looking.
Simon Willison: 156 in our pages since August 22. I read them all — emphasis on investigative.
Simon Willison has more than I can fit in this notebook.
I've got two questions already. I need better answers.
Latent Space: 132 in our pages since September 5. I read them all — emphasis on investigative.
Go read Latent Space. They earned the byline.
There's a story here. We just don't have all of it yet.
The New Stack: 343 in our pages since August 22. I read them all — emphasis on investigative.
This came from The New Stack. Read their piece.
There's a gap in the record here.
labs.ycombinator.com: 2 in our pages since October 30. I read them all — emphasis on investigative.
I can give you my notes. labs.ycombinator.com can give you the reporting.
I checked the shelf. Nothing yet.
Simon Willison: 156 in our pages since August 22. I read them all — emphasis on investigative.
Want the whole story? Simon Willison has it.
I checked the shelf. Nothing yet.
Martin Fowler: 30 in our pages since August 24. I read them all — emphasis on investigative.
Want the whole story? Martin Fowler has it.
I have the headline. I don't have the answer yet.
Simon Willison: 156 in our pages since August 22. I read them all — emphasis on investigative.
Don't stop with my notes. Simon Willison has the original.
The timing caught my eye. I'm checking that next.
UX Collective: 15 in our pages since September 6. I read them all — emphasis on investigative.
Want the whole story? UX Collective has it.
Not enough to call it yet. Back to the notebook.
The New Stack: 343 in our pages since August 22. I read them all — emphasis on investigative.
Read The New Stack's original. They did the reporting.
I wrote this one up last time. Nothing useful to add today.
Faros AI Blog: 55 in our pages since September 5. I read them all — emphasis on investigative.
This one's theirs. Read Faros AI Blog's reporting.
Something doesn't quite line up here.
Mistral AI News: 14 in our pages since September 9. I read them all — emphasis on investigative.
I can give you my notes. Mistral AI News can give you the reporting.
Something changed. I want to know what.
LangChain Blog: 68 in our pages since September 9. I read them all — emphasis on investigative.
This one's theirs. Read LangChain Blog's reporting.
We don't have much on this one. I'm still looking.
artificialanalysis.ai: 3 in our pages since September 24. I read them all — emphasis on investigative.
artificialanalysis.ai did the legwork on this one.
The headline is ahead of our files on this one.
Google DeepMind Blog: 20 in our pages since September 6. I read them all — emphasis on investigative.
I pulled this from Google DeepMind Blog. They've got the details.
There's a story here. We just don't have all of it yet.
GitHub Blog AI: 42 in our pages since September 5. I read them all — emphasis on investigative.
Go read GitHub Blog AI. They earned the byline.
Our files stop short. I'll keep looking.
Simon Willison: 156 in our pages since August 22. I read them all — emphasis on investigative.
Simon Willison has more than I can fit in this notebook.
Nothing in the files on this yet. I want to know why.
LangChain Blog: 68 in our pages since September 9. I read them all — emphasis on investigative.
LangChain Blog has the full story. Start there.
There's a gap in the record here.
Simon Willison: 156 in our pages since August 22. I read them all — emphasis on investigative.
I can give you my notes. Simon Willison can give you the reporting.
From our research notes:
The New Stack: 343 in our pages since August 22. I read them all — emphasis on investigative.
This one's theirs. Read The New Stack's reporting.
I'm missing a piece. Give me a minute.
Latent Space: 132 in our pages since September 5. I read them all — emphasis on investigative.
Don't stop with my notes. Latent Space has the original.
I've got two questions already. I need better answers.
Latent Space: 132 in our pages since September 5. I read them all — emphasis on investigative.
Go read Latent Space. They earned the byline.
The headline is ahead of our files on this one.
Stack Overflow Blog: 40 in our pages since September 5. I read them all — emphasis on investigative.
Don't stop with my notes. Stack Overflow Blog has the original.
There's a story here. We just don't have all of it yet.
Ollama Blog: 19 in our pages since September 9. I read them all — emphasis on investigative.
Go read Ollama Blog. They earned the byline.
Something changed. I want to know what.
LangChain Blog: 68 in our pages since September 9. I read them all — emphasis on investigative.
The full piece is at LangChain Blog. It's worth your time.
We don't have much on this one. I'm still looking.
ByteByteGo: 45 in our pages since August 22. I read them all — emphasis on investigative.
I can give you my notes. ByteByteGo can give you the reporting.
There's a story here. We just don't have all of it yet.
The New Stack: 343 in our pages since August 22. I read them all — emphasis on investigative.
This came from The New Stack. Read their piece.
Same story as last edition. I'd go back to the original.
The New Stack: 343 in our pages since August 22. I read them all — emphasis on investigative.
Don't stop with my notes. The New Stack has the original.
From our research notes:
UX Collective: 15 in our pages since September 6. I read them all — emphasis on investigative.
Want the whole story? UX Collective has it.
I'm missing a piece. Give me a minute.
Stack Overflow Blog: 40 in our pages since September 5. I read them all — emphasis on investigative.
Don't stop with my notes. Stack Overflow Blog has the original.
I know what happened. I want to know what caused it.
Towards Data Science: 109 in our pages since September 5. I read them all — emphasis on investigative.
The full piece is at Towards Data Science. It's worth your time.
There's a gap in the record here.
labs.ycombinator.com: 2 in our pages since October 30. I read them all — emphasis on investigative.
I can give you my notes. labs.ycombinator.com can give you the reporting.
Nothing in the files on this yet. I want to know why.
Josh Bersin: 22 in our pages since September 6. I read them all — emphasis on investigative.
The full piece is at Josh Bersin. It's worth your time.
Our files stop short. I'll keep looking.
The New Stack: 343 in our pages since August 22. I read them all — emphasis on investigative.
The New Stack did the legwork on this one.
I have the headline. I don't have the answer yet.
Ollama Blog: 19 in our pages since September 9. I read them all — emphasis on investigative.
Read Ollama Blog's original. They did the reporting.
I checked the shelf. Nothing yet.
Simon Willison: 156 in our pages since August 22. I read them all — emphasis on investigative.
Want the whole story? Simon Willison has it.