We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Shipping agent code without reading all of it
Two practitioners explained this period how they decide what agent-written code a human actually reviews.
DHH (Omarchy). In a live demo of Omarchy 4.5, DHH said he hasn't written "a single line of serious production code" for Omarchy by hand since the Quattro release. Development has been agent-driven ever since . His loop: tell the agent what he wants, then keep tweaking and giving feedback "until everything looked just right." Building a traditional desktop mode on top of Hyprland took about a day this way . His review rule:
- Changes that are critical or touch userland get reviewed.
- Work that can be checked by a program, such as keeping up the test suite, runs through automated agent updates and "is not even fully human reviewed" .
- He says he adds little over "a pair of frontier agents" at judging whether code is correct. What he adds is "taste, direction, and scope" .
His tip for building new apps: point the agent at something that already exists and say "I want it in that structure" . A PDF-signing app came from one prompt ("something like Apple PDF signing where it can save the signatures"). It had no layout instructions and needed only one follow-up fix . A translation PR that Muse Wonder produced in a day added 350,000 lines. He notes they are not 350,000 lines of code, and the result is Omarchy in 50 languages .
Poteto. In slides for an upcoming AI Engineer talk, Poteto says his team ships "30x'd" amounts of code "without reading it," and performance is as good as or better than at launch. He credits "agent friendly architecture bets" with keeping slop under control. About 230 people landed at least one PR to Grok bot in October .
Coding-agent tool makers acquired
Ben Tossell listed three deals on one day: Humanlayer → Factory, Conductor → Meta, Pierre → OpenAI .
- Pierre → OpenAI: Greg Brockman says Pierre's open-source developer tools "already power parts of codex today" .
- Conductor → Meta: the team will build "the next interface" with Meta Superintelligence Labs (MSL) .
- Cat Wu → OpenAI: after leaving Anthropic, where she worked on Claude Code, Cat Wu is joining OpenAI to lead product management for Core Products .
New open-weight coding models
- SWE-2 Flash (Cognition): a fast coding model that Cognition says beats all American open-weight models on leading evals, including models 3x its size. Cognition isn't releasing the weights openly. Instead it is building a program to share them securely with partners and researchers . swyx says it is 3x faster and 5x cheaper with the same score on FrontierCode 2 .
- Ripple-65B Preview (RicursiveAI): a 65B dense model pretrained from scratch. RicursiveAI claims a new state of the art among US open-weight models on agentic coding, beating much larger models such as Inkling-975B and Nemotron 3 Ultra-550B . It says the full R&D cost was $32M .
OpenAI's agent pipeline for finding and patching vulnerabilities
OpenAI built an autonomous pipeline that discovers vulnerabilities, exploits them and writes patches. In one run with an upcoming cyber model, it validated and generated patches for 948 Linux kernel local privilege escalation (LPE) vulnerabilities, at a total API cost of $1.2M . Brockman puts that at about $1,000 per vulnerability . OpenAI's Eric Wallace says security researchers used to find about 20 Linux LPEs in a typical year. He says the model found about 1,000 in one week and discovered at least 10 new classes of exploitation technique. OpenAI is putting "a significant amount" of frontier compute into patching open source for free . Jarred Sumner predicts that "dependencies are over" because rewriting one is "one prompt away" .
Tool updates
-
Codex
/goal: local threads used to use more and more memory as they got long. Brent Traut says the leaks in long-running turns are fixed, so "you should be able to /goal as long as you like now" . -
Browser use in LangChain Managed Deep Agents: add a
browser.pyfile that callsdefine_browser. Setscope="thread"to give each thread a fresh browser, which suits end-to-end QA. Setscope="user"to share cookies and logins across threads . The agent gets four Stagehand tools: snapshot, run (code-mode actions), screenshot, and handoff, which passes the browser to you to sign in . - Sourcegraph Agentic Batch Changes can now connect to MCP servers such as Linear, Confluence and Buildkite .
-
Antigravity now renders
.ipynbnotebooks natively . A DeepMind engineer has also added T3 Code as a recognized client in Antigravity's ACP server (PR) . - T3 Code: Theo claims users "more than 3x their token throughput" after switching .
Smaller techniques
- Skip duplicate CI after a squash merge (Kent C. Dodds): when PR CI passes, record the Git tree hash (not the commit SHA) along with the workflow inputs. If the squash commit on main has the same tree hash, skip tests, lint and typecheck before deploying. If not, run the full CI suite .
- Clone a SaaS product instead of buying it (swyx, Latent Space): swyx offered $10,000 for an open-source clone of a $40,000 SaaS product and covered entrants' coding-agent subscriptions. He got 69 submissions, about five he would pay for, all within a week . One entry replaced clicking through 1,000 talk submissions one at a time with a single sortable table . The AI judge only checked whether entries were valid. Picking the best stayed with humans, because "I cannot make an AI judge for things that I don't know what I want" .
- Cost controls for agent tool calls (treg tutorial, shared by Jason Zhou): give each agent its own limited token, kept in its environment and never in a prompt. Try a tool on 20 rows and check the output by hand before automating, and log every call . Cap each paid tool before it goes into a loop . Use free sources first and a gate before each paid step, and run long chains on something that stays awake: closing a laptop paused a 21-hour run for 16.7 hours .
- The linked tutorial reports a research agent making 68,413 tool calls for $53.49. It presents treg as a single-token catalog and router with 4,100+ tools across 124 providers, searchable by job and compared by measured success rate, speed, and price; it connects to Claude Code, Codex, and Cursor via CLI, or hosted Claude and ChatGPT via MCP.
- For cost control, connect keys you already pay for and issue a restricted agent token kept in its environment, not prompts or shared files. Cap each paid tool before putting it in a loop—the author says Hunter spent $4.66 of a $4.85 balance in two days before he capped it. First test on 20 rows by hand, then log each call’s tool, cost, and result.
- A reusable pipeline pattern is free sources first, a cheap yes/no gate before each paid step, then auditing by cost per usable result; the author’s auditor identified five tools that cost $11.49 (21% of spend) without earning a place in his niche. Chain steps only after they work independently, and run long jobs on an always-on server or cloud agent; closing a laptop paused the author’s 21-hour run for 16.7 hours.
- Caveats: a prepaid balance can stop abruptly, and direct calls to one provider do not fail over; only routed tools do.
- DHH says Omarchy development became agent-driven after Quattro: for about three months and thereafter, he wrote essentially no serious production code by hand. His loop was to describe what he wanted, then tweak and give the agent feedback until the result looked right.
- For agent-built applications, he recommends pointing the agent to an existing app and asking for a similar structure. His PDF-signing app began with a broad request for reusable signatures like Apple’s; he gave no layout instructions and needed just one follow-up to make inserted fields appear at the cursor.
- He used Muse Wonder for Omarchy translations into 50 languages; he described the resulting change as 350,000 lines, clarifying that it was not 350,000 lines of code. DHH says he does not review every generated line: critical or user-facing changes get review, while agent-maintained tests and other programmatically verifiable changes may not get full human review; he sees his main contribution as taste, direction, and scope.
- Wang replaced a proposed $40,000 conference SaaS purchase with a challenge offering $10,000 for an open-source clone, with coding agents available for further customization; 69 submissions came in, about five were usable enough to pay for, and the work was done in a week.
- One implementation replaced click-by-click review of 1,000 submissions for 25 seats with a sortable table view. The AI judge screened for valid submissions rather than choosing the best; the team evaluated usefulness themselves, over roughly two months of testing across organizer, speaker, attendee, sponsor, and vendor roles.
- A conference operator replaced a proposed $40,000 SaaS purchase with an open-source clone competition: offered $10,000 for a clone, covered contestants’ coding-agent subscriptions, and received 69 submissions. About five were worth paying for, and the work was done in a week. The plan was to customize the result with coding agents; the operator accepted some downtime in exchange for software tailored to the business.
- A concrete customization improved a conference workflow: selecting 25 speakers from 1,000 submissions had required clicking through records one by one; a contestant replaced that with a sortable single-table view that the existing SaaS lacked and the vendor would have taken months to deliver.
- Separate eligibility checks from product judgment: the AI judge checked whether submissions were valid, not which was best; the team chose the most useful options after reviewing examples. Evaluating the conference software took about two months because five user roles needed to be tested with separate accounts and data entry.
RicursiveAI released Ripple-65B Preview, a 65B dense LLM pretrained from scratch, and claims it sets a U.S. open-weight state of the art on agentic-coding tasks, outperforming much larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B. Swyx relays RicursiveAI’s claim that developing and training the model cost $32M—one-tenth the cost of the prior open-weight agentic-coding state of the art; the figure covers full R&D from scratch, not an inference price.
Cognition introduced SWE-2 Flash as a fast, efficient coding model, claiming it outperforms all American open-weight models on leading evaluations, including models 3× larger; it said it is developing secure weight-sharing for partners and researchers to balance transparency with cyber risks. Swyx reports 3× faster inference and 5× lower cost with the same FrontierCode 2 evaluation result, and characterizes the approach as distillation without tradeoffs.
datasette-llm 0.1a10 is an LLM integration plugin for other plugins to depend on; it fixes a bug that caused the agent.datasette.io demo to expose models beyond those allowed by its configuration .
After trying Opus 5.5 on Claude’s site, Theo called it “miserable” as a personal-assistant model; this is a task-specific assistant assessment, not evidence about coding performance. In a related exchange, he contrasted a model that would not recommend T3 Code with ChatGPT, which he said would recommend it aggressively; he initially suspected the model had hallucinated things he said, then clarified that the behavior was not due to memory and that the model explained its reason in the next message.
Ben Tossell flagged three acquisitions: Humanlayer → Factory, Conductor → Meta, and Pierre → OpenAI.
Claude Code 2.1.317 adjusted its Opus 5.5 prompt after a rule against recreating specific games also blocked requests that cited them only to describe mechanics: users can now name a game as inspiration for an original game, while requests to copy it get an offer for something similar instead.
Kent Dodds highlighted Poteto’s claim from an upcoming AI Engineer talk: their team ships 30× as much code without reading it, with performance as good as or better than at launch. Poteto credits agent-friendly architecture bets with helping tame “slop” while maintaining a high bar for polish and shipping speed. Poteto also said about 230 people landed at least one PR to Grok bot in October.
Reacting to a report that Apple is cutting employee AI spending limits, Peter Steinberger questions individual limits and favors agents representing projects, running loops, and evolving codebases, with humans curating the automations as agent capabilities improve.
For Omarchy QA, ThePrimeagen says Dell supplied a workstation with an NVIDIA RTX 6000 Blackwell, Intel Core Ultra 9 285K, and 128 GB of RAM; he expects it to run local computer-use and decision models, and says it supports 30 QEMU instances or 60 Bun programs running custom test harnesses. He plans to use local models to generate dynamic tests and validate every PR and nightly run, potentially at no cost; he is still seeking recommendations for computer-use models.
Kent C. Dodds explored a CI optimization with ChatGPT’s “Cole · Coder”: record the Git tree hash and workflow inputs when PR CI passes, then compare them with the squash commit on main. If they match, skip duplicate tests, linting, and typechecking before deployment; if they differ, run the full CI suite. The proposed check uses the Git tree hash rather than the commit SHA.
- Browser access lets Managed Deep Agents work with sites lacking agent-native APIs, internal tools, and browser-based UI end-to-end tests; the setup requires a browser (a sandbox alongside the agent or a provider such as Browserbase) plus browser tools.
- The LangChain presenter says testing seven or eight tool sets led them to choose Browserbase Stagehand, which performed better in their tests, particularly on cost and latency. Their four tools are snapshot (interactive elements), run (code-mode browser actions such as clicking and scrolling), screenshot, and handoff so a user can take control and return the browser to the agent.
-
To enable browser use in Managed Deep Agents, add
browser.pyand calldefine_browser. Choose thread scope for a fresh browser and isolated login state per thread—useful for repeatable QA—or user scope to share cookies and logins across threads for a personal assistant; the demo also shows the agent handing the browser to the user for sign-in before resuming.
Armin Ronacher endorsed a prediction that dependencies may become less attractive when developers can rewrite them with a prompt. The prediction cites a Brotli vulnerability said to have affected .NET, Node, Bun, Envoy, and others, and asks why take dependency risk if a rewrite is “one prompt away.” This is a strategic forecast, not a reported workflow or tool release.
Theo claims T3 Code is “way better for developers” than Claude Code or Codex; the post text offers no concrete differentiators or workflow details and points to a video for the long answer.
In a Managed Deep Agents demo, Sydney Runkle used an agent with browser access to book dinner reservations on a site without agent-native APIs; after logging in once, the agent saved the login and acted on the user’s behalf in later sessions.
Theo reports that users moving to T3 Code have more than tripled their token throughput, which he credits to the app helping them use their AI subscriptions more fully; this is a throughput claim, not evidence of a corresponding increase in coding output. One user said they were using T3 Code more than they had ever used Claude or Codex.
Antigravity’s ACP server internally recognizes t3-code as a client; once T3 runs the upcoming release, Antigravity is expected to attribute traffic to T3 Code. A PR was made to switch T3 Code to that client identity; the post does not say the PR has merged or the release is live yet.
