We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Give agents a pass/fail check, then remove what slows it down
Addy Osmani frames running agents as managing people. His main rule: give the agent something that returns a definite pass or fail (tests, build exit codes, linters, diff scripts, screenshots) and have it show the result. Don't accept "looks done" . He notes the command that runs these checks is "the single most useful line" in CLAUDE.md . Ways to wire it into Claude Code:
- Ask: "Run the auth tests after, and continue until they pass."
-
/goal: a separate model checks a condition after every turn. - Stop hook: blocks the turn from finishing until, for example, the typecheck passes.
-
/code-review: sends the diff to a fresh subagent. Scope it to correctness and tell it to ignore style, or it will always find something .
His other rules:
- Correct twice, then write it down. Put it in CLAUDE.md or a shared skill .
- Use hooks or deny rules only for hard requirements. Drop ALL-CAPS instructions, "think harder" and "don't be lazy" .
-
Plan, then split the work. Agree a plan first. Use
/batchto split a mechanical change into 5–30 worktree tasks, each with its own PR. Check the slices don't overlap and land shared interfaces first . - Reset after two failed corrections. Start a fresh session with a better prompt .
- Answer five questions before an unattended run: What does done look like? How does it verify itself? What are its boundaries? What turn or spend limits apply? Where does output go, and who approves it? Pilot on three files first .
Theo's new video ("Coding Isn't About Code Anymore") explains why this matters more for agents than for people. A 5-minute dev-server start blocks an agent "insanely deeply" and, with 5-minute cache TTLs, wastes tokens on missed cache reads. Slow CI limits how much work you can run in parallel . His examples of this upkeep:
- He deleted about half of T3 Code's unit tests, roughly 250,000 lines. He restored 4 of 11,000 and reports no regressions the deleted tests would have caught .
- When a bad test reappears, the contributor goes back to the agent thread and asks the agent why it added the test. The team then fixes the root cause .
- A "thread transfer impact" check runs on every PR. It replays real Codex and Claude thread data through a simulated pipeline and flags data-transfer regressions in the PR, where a babysitting agent will see them. It takes under 45 seconds .
This adds to last period's verify-before-review material. The check is part of the repo itself, and it tells agents what to optimize for.
Long AI-written PR descriptions: write a short human intro, collapse the rest
Simon Willison treats PRs opened by Claude Code or Codex as his unit of work . To keep the agent's description readable, he opens with a short paragraph he has vetted, sometimes adds a screenshot, and puts the full agent-written text inside <details><summary>. GitHub Markdown supports this. The summary line is the model name and effort level, such as "Claude Opus 5.5 Medium" .
He also shared a prompt sequence worth copying. He asked Opus 5.5 in regular Claude chat to compare a SQLite parse_number() extension written in C and in Rust. The Rust builds were 298–321 KB against 14 KB for C, but they were faster . After he asked it to write a custom C decimal parser, the new C version ran at 16–19 ns per row, against 32–55 ns for the Rust versions. He then had it package the work as a report in simonw/research .
Don't trust agents on MCP auth
Theo says that every time he has Opus set up an MCP authentication path, it "makes horrible assumptions" and the result doesn't work . Asked about remote machines, it proposed curling a script, copying codes by hand on mobile and notifying the agent manually . He is writing an "MCP the right way" guide for his agents. Their first draft rendered the MCP URL as a clickable link, and he warns not to trust anything agents say about MCP or connection strategies . If you build MCP auth with an agent, test the remote case yourself.
On usage, Theo ran 8–25 Opus 5.5 threads most of the day and used less than 20% of a $200 Claude plan . His Codex plans get the heavy use because he reviews with Sol, often while Opus reviews the same code in parallel . He also banned a user whose bot left more than 1,000 "review" comments on a T3 Code PR .
Mihail Eric: research the code first, and understand your harness
On Latent Space, the Stanford CS146S instructor argues that large codebases need more than a single prompt:
- research the code
- propose alternatives for a human to choose between
- write a plan/spec
- implement and test
He teaches the harness itself. Students build an agent loop of about 200 lines, then proxy Claude Code's traffic to read its system, plan and compaction prompts . swyx says context for subagents is still "very hacky": they over-fit to stray comments when given too much .
At Monaco, Monocoder (built on Coder) provides isolated environments for long-horizon multi-agent work, and a Slack agent named HAL debugs issues and opens PRs . Their rule is "buy to learn, build to keep." They still use Devin for code review .
Claude Duet: how to start new Claude Code sessions
Riley Brown found that asking Duet to create a Claude Code session doesn't work. Use New Session on your phone, pick the computer and folder, and type the first prompt . Requests to existing sessions need approval and Face ID . He runs Amphetamine so his Mac never sleeps . He rates Claude Code's computer use well below Codex's .
A counterpoint on MaxxBench
Last period's MaxxBench result showed large score drops on modified DeepSWE tasks. Salvatore Sanfilippo (in Italian) argues those drops overstate how much weaker the models are. Some modified specs leave out details the hidden tests check. Models sometimes target the tests they expect rather than the written spec. Smaller models are less able to override what they learned earlier . He proposes about 100 genuinely new tasks of the same difficulty, run on the same harness .
Quick notes
-
DeepSeek V4.1 Flash 1017 Ultrafast is on Vercel AI Gateway. Vercel says it generates over 1,000 tokens per second; the model ID is
deepseek/deepseek-v4.1-flash-1017-ultrafast. - TensorFold recipes are in Omarchy's local-AI catalogue (omarchy-local-ai) for Qwen3.8-27B, Flash Next and Nemotron 3.5 Bolt . DHH says TensorFold will be key to Omarchy's out-of-the-box local agents on NVIDIA .
- Asahi Linux published an LLM policy that "broadly forbid[s]" generative AI . DHH says Asahi rejects upstreaming his group's agent-accelerated Apple Silicon work . Armin Ronacher defends Asahi's right to the policy .
- Open question: ThePrimeagen says tasks that took Claude 5.5 10 minutes take 40 as a project grows, and asked for ways to prevent it .
- To understand a coding agent rather than treat it as a black box, build a minimal agent loop (about 200 lines, without subagents or MCP), then proxy a full harness’s requests and responses to inspect its system, planning, and compaction prompts.
- For unfamiliar or large codebases, go beyond one-shot prompting: research the code, propose alternative approaches for a human to choose from, then write a plan/spec before implementation and testing. For multi-agent work, fan out exploration and subtasks deliberately, manage what context subagents inherit, and consolidate or refresh context; the right balance remains unsettled.
- Before scaling parallel or background agents, make the repository understandable and verifiable, with clear service boundaries and useful documentation; agent readiness is a foundation for scaling those systems. As code generation accelerates—one example raised is 30+ PRs a day—review becomes a bottleneck: decide what can be auto-approved versus what needs human scrutiny. Monaco’s detailed PR descriptions summarize the implementation and point to changed files and service boundaries.
- Treat skills as reusable procedures, but maintain them: derive a skill from repeated work and iterate on it in chat; watch for triggers that fire in unrelated tasks and prune unused skills. Skills can also encode organizational practices such as preferred APIs and service boundaries, rather than only personal shortcuts.
- At Monaco, Monocoder—built on Coder—provides isolated environments for multi-agent, long-horizon coding tasks; its Slack-based HAL agent helps debug issues, open PRs, and answer product questions. The team’s stated heuristic is “buy to learn, build to keep”: it retained Devon for code review after evaluation, while building internally where available tools lacked maturity or fit its infrastructure.
- Repository-based coding benchmarks typically give an agent an existing project, a specific commit, a task specification, and a minimal command-running harness; hidden fail-to-pass tests check the requested change, while pass-to-pass tests check that existing behavior still works.
- Sanfilippo cautions against treating a score drop on modified benchmark tasks as a direct measure of weaker general coding ability: variants may become harder or underspecify the hidden tests, and agents may follow what they expect the evaluator to want instead of the written specification. He argues that smaller models are more susceptible to such conflicts with learned priors than larger models.
- For a cleaner comparison, he proposes creating about 100 genuinely new tasks on real repositories, preserving the engineering challenge and difficulty while avoiding superficial traps, and using the same harness and scaffolding with some human review. Comparing models on these tasks would better distinguish benchmark familiarity from the ability to solve similar coding problems.
- Swyx said Cursor ended 2025 at $1B ARR and that its enterprise business was growing 300% quarter over quarter—a signal of the tool’s commercial momentum, not a workflow result.
- AI Engineer World’s Fair made its full Forward Deployed Engineering track available, including sessions with Cursor VP of FDE Pauline Brunet and Cognition deployed-engineering lead Jia Wu; the track playlist is a relevant practitioner resource.
- Claude Duet lets a phone connect to Claude Code sessions running on a local computer; in the demo, the creator asked an existing MacBook Pro session to restyle the Agent Native app and push it to production, then checked the live link from the phone.
- To start a new remote coding session, use Duet’s New Session flow to choose a computer and folder and send the first prompt; messaging Duet alone won’t create the session. The computer must stay on, and the creator uses Amphetamine to prevent their Mac from sleeping.
- Duet also supports voice requests to existing Claude Code sessions; the demo includes approving a request from the phone and checking that the session is working.
Simon Willison used regular Claude chat (Opus 5.5) to compare C and Rust implementations of a SQLite parse_number() extension, prompting it to assess performance, code complexity, and binary size; the initial Rust extensions were larger (298–321 KB vs. 14 KB for C) but faster. After identifying decimal parsing as a bottleneck, he asked Claude to write a faster C parser; the new C version measured 16–19 ns per row across mixed, integer-only, and decimal-only data, versus 32–55 ns for the Rust versions. He then asked Claude to clone his research repository and package the investigation as a report, turning the prototype work into reusable documentation.
-
Give each task a definitive self-check—such as tests, a build, a linter, or a screenshot—and put the exact validation command in
CLAUDE.md; ask the agent to run it and fix failures. Claude Code options include/goalfor condition checks during long work, a Stop hook that blocks completion until a check passes, and/code-reviewfor review by a fresh subagent. -
Brief the agent on what it cannot discover in the repository: the objective and why it matters, constraints, out-of-scope work, and success criteria. Agree on a plan and inspect its assumptions before coding; parallelize only non-overlapping work.
/batchcan divide a mechanical change into 5–30 tasks in separate worktrees and create a PR for each, but check for overlap and land shared interfaces first. -
Make repeated corrections compound by recording them in
CLAUDE.mdor a shared skill; after two failed attempts to steer the agent, start a fresh session with a better prompt. Keep product, security, and compliance decisions with a human: a green test suite is not approval, and the reviewer remains accountable for the diff. - Before an unattended run, define what “done” means, how it will self-verify, its boundaries and turn/spend limits, where it reports results, and who approves consequential actions; pilot on three files and watch the first runs before trusting it unattended.
- Tune the repository as infrastructure for agents: reduce dev-server startup time, CI time and resource use, and release friction. Theo says slow validation can waste tokens, while slow CI limits parallel work and slow releases deter frequent merges.
- Treat recurring agent mistakes as feedback about the codebase: when bad tests reappeared, the team asked contributors to find the agent conversation and ask why the test was added, then used the answer to address the root cause and adjust repository guidance. As an example, Theo says he removed about half of T3 Code’s unit tests—around 250,000 lines—and restored four of 11,000; he reports no regressions that the deleted tests would have caught.
- Add fast, agent-visible checks for behaviors you care about. T3 Chat’s per-PR transfer-impact action runs real thread data through a simulated end-to-end pipeline and flags meaningful data-transfer regressions; it gives agents and reviewers feedback on the target metric and takes under 45 seconds to run.
- For GTM research, treg is presented as an “OpenRouter for agent tools”: one token/balance can access multiple tools, while installable skills package instructions for complete jobs.
-
Setup for Claude Code, Codex, or Cursor: install treg with
curl -fsSL https://treg.to/install.sh | sh, log in, then add skills withnpx skills add superdesigndev/treg. Set a token cap and per-callX-Treg-Route-Max-Costceiling; test on five rows, inspect the output manually, and log each tool, cost, and result before automating. - Get each agent working on its own before chaining them; use one Google Sheet for agent handoffs, run cheaper steps before paid ones, and put long-running chains on an always-on server or cloud agent rather than a laptop.
- In the demonstrated setup, Claude Duet can show recent Claude Code sessions on connected Macs and route requests to a local session; from his phone, the creator asked a MacBook session to restyle an app and push it to production, then approved and authenticated the request.
- Starting fresh requires a different flow: use Duet’s New Session, choose the computer and folder, and enter the first prompt; simply asking Duet to create a session does not work. To keep remote access available, the creator leaves the Mac on and uses Amphetamine’s indefinite session to prevent sleep while it is plugged in.
ThePrimeagen reports that, as a project progresses, a task that “clause 5.5” previously handled in 10 minutes now takes 40, and asks for ways to prevent the slowdown.
A same-thread workflow: ask Claude Opus 5.5 Medium in regular Claude Chat to port a Three.js 3D project to Vanilla JS; Willison received a Vanilla WebGL version, though it differed from the Three.js version in ways he did not fully understand. The result suggests Opus 5.5 can produce 3D modeling in Vanilla JS + WebGL without Three.js; Willison found the Three.js code slightly easier to understand, but rarely examines code for these projects.
Willison uses Claude Code- and Codex-opened PRs as his unit of work, iterating on them until a feature is complete. To make agent-written descriptions easier to scan, he puts a short, human-judged explanation first and collapses the full description in GitHub-supported <details><summary> markup; he uses the model name, sometimes with effort level, as the summary and may add a screenshot.
ThePrimeagen describes an OS-install workflow: install QEMU and related utilities, download the Omarchy ISO, give “luna none” screenshots and keyboard input, and have it create an Omarchy account and reboot into the installation; he says this is easier than first starting a machine with Windows installed .
The speakers suggest company-private Q&A could still be useful with coding agents: agents might ask and answer questions in the Q&A system, or be given access to company repositories and asked to find an answer. They present these as possibilities, not as a reported deployed workflow.
Mihail Eric’s Stanford CS146S students practice agent-era software work by opening a PR to an open-source project and explaining their choices to the maintainer, who decides whether to merge it. At MonacoGTM, his team buys code review, builds its own agent workspaces, and gives its on-call agent enough access to investigate failures and open a PR; the approach is summarized as “buy to learn, and build to keep.”
Simon Willison hides the bulk of his AI-generated GitHub PR descriptions inside a collapsed <details><summary> block, which GitHub Markdown supports, to make lengthy AI-written descriptions less exhausting to read .
Vercel announced DeepSeek V4.1 Flash 1017 Ultrafast on AI Gateway, claiming it generates over 1,000 tokens per second and recommending it for coding, reasoning, and agents. ThePrimeagen said he had to try it, but shared no test or results.
- Theo ran 8–25 concurrent threads for much of the day, all using Opus 5.5, while using less than 20% of a $200 Claude subscription.
- For code reviews, he uses Sol and usually runs Opus reviews on the same code in parallel; he says this is why his Codex subscriptions are being heavily used. He says the activity reflects working to land changes, not current “death loops.”
Riley Brown reports using Opus 5.5 with a Blender video-editing workflow to make a launch video in under 10 minutes from a single prompt: create a Screen Studio-like video showing a desktop app and an iOS app. He then prompted it to make a vertical, UGC-style version; the post shows the request, not confirmation that the version was completed.
Theo reports that Opus repeatedly makes poor assumptions when implementing MCP authentication, resulting in flows that do not work. He gives a remote-machine example: browser sign-in is unavailable, while the fallback requires running a script to generate a secret, copying an obscure code on mobile, opening a link, and manually notifying the agent when complete. This is a concrete warning to test agent-built MCP auth flows against remote use and minimize manual handoffs.