We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Anthropic: humans design, Claude ports, shadow traffic verifies
On Latent Space, Anthropic's platform team said it has fully rolled out a Rust version of the Messages API, replacing a CPU-hungry Python monolith . When they chose Rust, the team had no Rust developers. Claude wrote the code, and humans designed the distributed-systems architecture . They said the two things that mattered most were the plan and the verification. They built a complete Rust fork of the API, ran shadow traffic through it, and compared its outputs with the Python system's . They ported the whole system and then turned it on gradually, instead of the piece-by-piece migration they said would normally take years . They also kept it a pure port and held back new features . They estimated about 40–50% CPU savings and a 10–20% improvement in time to first token .
The steps you can copy: write the plan yourself, have the agent do a scope-frozen port, and check it by comparing outputs from the old and new systems on real traffic.
Their other advice from the same episode:
- Evals: the most common mistake is having too few. Start simple, build a fairly large suite, and let agents help it grow . Treat engineering like research: define the goal as an eval, then check whether a failure comes from the eval, the environment or the model. "A lot of times your eval is wrong" .
- Multi-agent: use one lead and split work that can run in parallel across subagents that share context. The subagents should run asynchronously and not block each other. "Other than that, like let it rip" .
- Sandboxing: Claude Managed Agents keeps the harness (the "brain") on a durable server, separate from the sandbox (the "hands"). The agent sees only placeholder credentials. An egress proxy injects the real ones from a separate vault, and network access can be limited to an allowlist of domains or turned off .
LangChain open-sources its design agent harness
LangChain's Content Studio is a Slack agent that makes slides, charts and videos. It runs on Managed Deep Agents, and LangChain has published the harness for others to fork . LangChain reports median model spend per request fell from about $8 to about $0.50, and its own code shrank about 80% after moving to Managed Deep Agents . Several of the patterns apply directly to coding agents:
- Everything is code, in a persistent sandbox. Each Slack thread gets its own Linux sandbox. Slides are HTML, charts are data specs and videos are animation timelines, so an edit can change one value and rerender .
- Rules that must always hold go in code, not the prompt. The agent sometimes skipped a review step that was only described in its instructions. The pipeline now blocks assembly until every part passes review: "The order that must always hold lives in code, and the judgment stays in the prompt" .
- Skills are scoped to jobs. A router skill maps request types to specialist skills. Each of the 48 skills reads like a routing rule, including when not to use it . MCP tools sit behind search, and their schemas load only when needed .
-
The main agent acts as a "chief of staff." Subagents run asynchronously. Each gets a brief covering its role, context, which files it owns and how to validate its work, plus a model and effort level chosen for the task. The coordination tools are
start_task,check_tasks,cancel_task/resume_task,report_progressandschedule_wakeup. -
Memory stays small.
AGENTS.mdis limited to a 200-line index, after an earlier version grew to 500K tokens. After each job, a subagent records up to five evidence-backed lessons, and a weekly pass promotes the useful ones into skills . - Remove workarounds when models improve. After Opus 5.5, they removed most hard-coded rules and kept only principles. They got more variety and less repetition .
Harrison Chase's Interrupt London talk covers similar ground. LangChain's open-source OpenWiki builds a coding knowledge base from your repo. It can also read Codex or Claude Code traces to find what the agent struggles with . His eval advice: grade the final environment state or the tool-call sequence, not just the answer . Run evals that write data against a mock or dev environment, not production .
Claude Code mods: build them by describing them
Addy Osmani posted a guide to Claude Code mods. Mods are JS/TS files that react to session events, change Claude Code's behavior or draw custom UI . They are hooks chained like middleware. Each hook can observe an event, rewrite it, or answer it so the default behavior never runs . To make one, describe it to Claude, approve the files and enable hot reloading. Then keep asking for changes without restarting . Mods made this way last only for the session. To keep one, copy it from ~/.claude/dev-mods/<session-id>/ and install it as a plugin . The API may change, so use the generated types in .claude-plugin/types/ as the reference for your version .
Example code for Token Weather (context usage), Blast Radius (confirming risky commands) and Replay Theater (per-edit diffs, including subagent edits) is in anthropics/claude-code-playground . Mods are not sandboxed and run with your permissions . Blast Radius only reads the command text, so $(…), aliases and scripts can get past it. Use permission rules if you need a hard block .
Benchmarks: a coding test graded on whether it would be merged
Artificial Analysis Intelligence Index v5 adds FrontierCode 2.0, Cognition's private benchmark. It grades real-world coding tasks on whether a repository maintainer would merge the agent's change . On the overall index, Opus 5.5 (max) leads at 58, followed by GPT-6 Astra at 57 and GPT-6.1 Sol at 55. The best released open-weights models score 32 . Among the top 10, cost per task runs from $0.96 for GPT-6.1 Terra to $13.04 for Claude Fable 5.1 .
Short notes
- Cursor design mode added Annotate. You can edit elements, draw on the page and check different screen sizes as visual prompts for the agent . Open it with Cmd+Shift+D in the browser (docs) .
- Kent C. Dodds says he ships dozens of PRs a day with Cursor and rarely opens the Cursor app. He drives it through Grok Bot . When his Cursor credits ran low, he had the bot move work to Devin's cloud agent, with Kody handling the handoff . He says he would use the Claude and OpenAI cloud agents more if they had APIs . On dropping PRs, he says he would first have to rebuild what PRs give him: AI review, preview deploys, CI and a record of decisions .
- Matt Pocock trusts agents most on tasks where they have all the data, can run experiments, can verify results on their own, and can hill-climb toward the goal .
- Steve Yegge says his Wheelhouse software factory produces 2,000–4,000 commits a week and runs without sandboxing. Its constraints are doctrine and policy, human gates and 14 programmatic denials . He warns these setups are expensive, hard to maintain and risky in enterprises .
- T3 Code: Theo says a change from Cursor broke the latest nightly. He is building a pipeline to prevent this and recommends the stable channel .
- Anthropic’s team ported its Messages API from a Python monolith to Rust despite having no Rust developers at the outset; humans designed the distributed-systems architecture and Claude wrote the implementation. They built a full Rust fork, ran shadow traffic through both systems, and compared outputs before rollout; they kept the migration to a port rather than adding features. The team reported roughly 40–50% CPU savings and 10–20% improvement in time to first token.
- Build a real eval suite early: start simple, expand it over time, and use agents to help evolve it rather than hunting for a perfect small set of tests. Treat engineering like research: define the goal, formalize it as an eval, then check whether a failure comes from the eval, the environment, or the model.
- Claude Managed Agents offers programmatic APIs for defining an agent and starting a session, so developers can skip building much of the agent infrastructure and harness themselves. For custom harnesses, the guidance is to start with simple primitives and experiment with multi-agent strategies: use one lead, delegate parallelizable tasks with shared context, and keep subagents asynchronous and non-blocking.
- For safer agent execution, separate the durable harness and agent state (“brain”) from the sandbox that executes work (“hands”); keep credentials in a separate vault and inject them through an egress proxy, while restricting network access with domain allowlists or disabling it.
- For Claude API integrations, use the Claude API skill in Claude Code or another coding agent, or provide a reference architecture for the agent to implement; the interviewees describe these as useful alternatives to relying on a large, complex cookbook when building different implementations.
- LangChain's Content Studio is an open-source Slack design agent built on Managed Deep Agents. Each Slack thread gets an isolated Linux sandbox that preserves its project; code-based assets and a shared project folder let the agent and subagents revise and rerender work without starting over. Teams can fork the repo, add brand tokens, fonts, and skills, and deploy it to Slack.
- Separate agent context by function: a router maps request types to workflows and specialist skills; 48 job-scoped skills span nearly 1,000 files, while a wiki centralizes company and design references and code gates enforce measurable checks. To limit context use, expose MCP tools through keyword search and load their schemas as needed.
- Encode subjective judgment as design principles, guardrails against common AI patterns, and the reasoning behind feedback. Their fast learning loop saves feedback to shared or personal memory files; after each job, a subagent captures up to five evidence-backed lessons, with useful lessons incorporated into skills and the wiki weekly. After Opus 5.5 improved output, they removed most hard-coded rules while retaining principles and high-level guidance.
- For expensive media work, explore in a storyboard first—about 50 ms per frame versus 8–40 seconds to render one second of finished film—then approve a direction before animatic and production; enforce this order in code because prompt instructions alone sometimes let the agent skip review.
- For hours-long projects, make the main agent a coordinator that delegates asynchronous tasks with briefs specifying role, context, file ownership, and validation, and selects model and effort based on task complexity. Keep the team able to steer the work with progress updates, a plan and ETA, live status and spend, and an approval step before publication. In an overnight 52-slide reskin, the agent reviewed three sample slides before building the rest in approval batches; the main agent ran for about an hour while subagents worked for more than six hours.
- LangChain reports that Content Studio has delivered over 3,000 assets, including more than 80 decks; median model spend per request fell from about $8 in mid-September to about 50 cents, and its code shrank about 80% after migrating to Managed Deep Agents.
- Coding-agent memory: LangChain’s open-source OpenWiki is a coding-focused knowledge base populated from a repository; integrations inspect Codex and Claude Code traces for what the agents struggle with and feed those findings back into the wiki.
- Improvement loop: Preserve exact inputs, outputs, and event sequence in traces; inspect failures, add them as eval cases, then propose changes to the model, harness, or context and use evals to check for regressions.
- Evaluate actions, not just answers: Set criteria for the final response, resulting file or environment state, or tool-call sequence. Run write-affecting end-to-end evals in a mock or development environment rather than production.
- For a bespoke PDF reader, the presenter used Opus 5.5 to draft a detailed build prompt, then gave it to DeepSeek at high-thinking settings, specifying Electron, React, TypeScript, and PDF.js. The resulting Folio included PDF reading and annotation features such as highlights, drawings, notes, signatures, dark mode, export, and printing; he says he built it locally, or that inference would cost less than a couple of dollars.
- In a local-model trial, he ran DeepSeek on a DGX station and had it build GridLine, an offline-first browser spreadsheet, in 50 minutes; he says a quantized version can run on a decent high-end consumer GPU. GridLine handled formulas, fill-down, and charts, but did not have all of Excel’s functionality.
- For a capability test, he deliberately used the simple prompt “Build a 3D simulation of a Rubik’s cube. Scramble and solve” to see whether the model could infer the expected details without a highly explicit specification; he reports that the scramble and solve worked accurately.
-
To build a Claude Code mod, describe the behavior or UI you want, approve Claude’s generated files, and enable hot reloading; then refine it with follow-up requests without restarting. The API can change between releases, so use the type declarations generated in
.claude-plugin/types/as the reference for your build. A mod made this way is session-only; copy it from~/.claude/dev-mods/<session-id>/and install it as a plugin to keep it. - Mods use event hooks chained like middleware: hooks can observe events, rewrite them for later hooks, or handle them instead of Claude Code’s default behavior. Practical examples include Token Weather for context-window usage, Blast Radius for previewing and confirming risky commands, and Replay Theater for reviewing turn-by-turn edit diffs, including subagent edits.
-
Treat mods as trusted local code, not a security boundary: they are unsandboxed and run with your permissions, while Blast Radius can miss commands involving substitutions, aliases, or scripts. Inspect the source and run
claude plugin validate; use permission rules for hard blocks.
An “Apple II Agent” using Opus 5.5 drew a pelican riding a bicycle on an Apple II, with examples shown in color and monochrome . Simon Willison praised the results .
- Keep creative judgment human: Tossell decided to manually identify which agent sessions could become useful stories, and to clarify each repeated workflow before automating it. He uses agents for execution-heavy setup such as preparing recordings and exports, editing transitions, uploading videos, and publishing pages.
- For visual explanations, he asked an agent to combine TL Draw with Pelle, which encodes layouts in code so visual styles can be applied. He steered the result toward an HTML site rather than a canvas-only output and had the agent restore missing annotations and arrows. He also explored having the agent build reusable visual components, such as consistent boxes, agent icons, webpages, people, and cameras.
A Latent Space episode featuring Anthropic’s engineering and product leads covers how Anthropic builds Claude with Claude; its platform is used to build first-party harnesses including Code, Tag, and Duet, and by more than 50% of Anthropic’s top customers.
-
Give the agent a persistent, editable workspace. Content Studio creates an isolated Linux sandbox for each Slack thread, where the project persists and subagents share files. Because outputs are code, an edit can change one timing value and rerender instead of rebuilding an asset; its motion toolkit includes HyperFrames for linting, frame checks and rendering, plus
motion-studiofor original per-pixel animation . - Keep task context modular and load tools on demand. A router maps request types to job-specific skills; the 48 skills (about 1,000 files) compose rather than duplicate craft, while a central wiki holds shared brand and product context. MCP tool schemas are searched and loaded as needed to limit context use .
- Retire workarounds as models improve. The initial harness supplied 115 reusable components, but the agent repeated layouts; after a reported quality jump with Opus 5.5, the team removed most hard-coded rules and kept high-level principles, reporting more variety and less repetition. They handle measurable issues with deterministic code checks and capture designers’ reasoning for subjective judgments in skills .
- Make expensive work spec-driven. A storyboard frame takes about 50 ms to render, versus 8–40 seconds for a second of finished film, so the agent explores and gets approval on a storyboard before creating an animatic and entering production. Since instructions alone sometimes failed to prevent skipped reviews, code now blocks assembly until every act passes review .
- Run long jobs asynchronously with a steerable coordinator. The main agent divides work into background tasks with a role, context, file ownership and validation criteria, choosing a model and effort level for each task. The Slack thread reports plan, ETA, live status and spend, acknowledges edits and requests approval before anything goes public; people can steer the agent while work continues .
-
Build feedback into memory loops. The agent saves preferences in the same turn to shared-team or personal memory; after a job, a lightweight subagent records up to five evidence-backed lessons, which later jobs rate and a weekly process promotes into skills or the wiki. Their
AGENTS.mdfiles are limited to a 200-line index after an earlier version grew to 500K tokens .
The team open-sourced the harness, including its sandbox, skills, wiki, memory, async subagents and Slack UX, for others to fork and customize .
ThePrimeagen is considering trying the linked GrokBot multi-agent setup and asked whether followers want a video about his experience; he has not reported a test or results yet. The linked post pitches a team of 50+ agents organized as a Chief of Staff, three PMs, and 40+ workers, but that is the other poster’s description, not ThePrimeagen’s firsthand experience.
Steve Yegge says his Wheelhouse software factory, running on a light multi-machine setup, produces 2,000–4,000 commits per week; its agents work around the clock on building, reviewing, and deploying features, as well as other startup operations. He favors an open, exploratory setup over sandboxing: agents can see the whole system and act independently, with doctrine and policy, human gates, and 14 programmatic denials as constraints; they work under his identity while signing their work under their own names. He cautions that software factories can be expensive, hard to learn and maintain, and risky in enterprises—and that tighter restrictions can undermine the collaboration and exploration that make them productive.
Steve Francia worked with Gemini to get a full NES emulator running on a GitHub Universe conference badge, then deployed Game Boy and Game Boy Color emulators on it; with the ROM, the badge can run The Legend of Zelda: Link’s Awakening .
Claude Code’s quoted release note says --dangerously-skip-permissions still blocks some obvious commands, such as rm -rf ~, but setting CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1 can disable those remaining blocks. The post responds with an emphatic warning: “DO NOT USE OR YOU WILL BE FIRED.”
Kent C. Dodds said eight hours had passed since an agent session that ran past his bedtime; he shared no workflow details. PowerSync sponsored the accompanying video and was promoted for real-time and offline-first capabilities for users and agents, not described as a tool Dodds had personally used.
Kent C. Dodds says he is adding an “execution contexts” primitive to Kody, which he says will enable new things with agents; he provides no implementation details or availability information.
Kent says he uses Cursor for most of his work, ships dozens of PRs daily, and rarely opens the Cursor app or site . Asked how, he says he uses Grok Bot (@bot), not t3 code or the CLI .
Matt Pocock describes agent-friendly tasks as ones where the agent has all the data, can run experiments, verify results independently, and “hill-climb” toward the goal; he says trust and results are strongest on these tasks. Kent C. Dodds replied that this category may soon expand to “everything.”
A developer integrating Kody into their workflow says it could be a major unlock for cross-agent work and that its secret management solves a real pain point, though it took time to understand the tool . Kent C. Dodds says he wants to help users reach that “light bulb moment” faster .
- LangChain’s open-source Content Studio is a Slack design agent built on Managed Deep Agents. Each Slack thread gets an isolated Linux sandbox that preserves the project across turns; assets are built as code, so the agent can make targeted edits and rerender, and subagents share the project folder.
- For complex jobs, the main agent acts as a coordinator: it delegates isolated tasks to async subagents, gives each a brief with its role, context, file ownership, and validation criteria, and selects a model and effort level for the task. People can steer work mid-run; Slack shows the plan, ETA, progress, spend, and an approval step before anything goes public. An overnight reskin of a 52-slide deck used three sample slides and approval before the agent continued in batches; the main agent ran about an hour while subagents worked for more than six hours collectively.
- Keep knowledge and execution rules structured: job-specific skills route work, a shared wiki holds reusable context, and deterministic checks handle measurable requirements. The team also enforces required workflow stages in code while leaving creative judgment to instructions.
- Treat model-specific workarounds as temporary: after Opus 5.5 improved, LangChain removed most hard-coded rules while retaining design principles, and reported more varied, less repetitive results. Content Studio has delivered over 3,000 assets; median model spend per request fell from about $8 to about $0.50, and the team’s code shrank about 80% after migrating to Managed Deep Agents.
Kent C. Dodds says he has repeatedly made over 100 releases a day as a solo maintainer and could work without PRs, but would need to recreate the confidence-building functions they provide: AI review, preview deploys, CI, and historical decision records.