🔥 TOP SIGNAL

Claude Code Projects is the clearest productization yet of the long-running coding-agent harness. Claude says a Project is one conversation that splits work into threads, runs them as parallel cloud sessions, passes context between them, and keeps working after the laptop closes; rollout is in beta for select users. @ClaudeDevs Boris Cherny says it changed how he codes: he stopped managing sessions, sends thoughts as they arrive, and lets the project remember how he works. @bcherny The practical alpha is goal-level delegation paired with proof: Theo’s T3 Code loop starts from a screenshot and ends with a tested PR plus a video, not a chat transcript. Theo - t3․gg

⚡ TRY THIS

• Turn a Project into your coordinator. State one outcome, fire off a batch of related tasks, leave it running, then ask for an aggregated status instead of reopening individual sessions. Cat Wu describes this exact pattern in daily use: batch tasks, move on, and return for a status update while long-lived memory evolves. @ClaudeDevs +1

• Use a proof-carrying bug prompt. Paste the screenshot and ask: Fix this. Test it. Record a video showing it works now and link me the PR when you're done. Babysit it until all the issues that come up in review are addressed. Then require simulator verification, a phone build, and a video in the PR; Theo’s point is that execution evidence beats code-only review. Theo - t3․gg

• Optimize for the floor, not the demo. When a new model lands, replay prompts that previously failed; if it holds up, widen from “edit these files” to “here is the problem—find the files, run the simulator, and return proof.” Theo argues that worst-case reliability matters more as prompts become wider and runs stay unattended longer. Theo - t3․gg

• Treat compaction summaries as untrusted input. Add a long-run test that captures the summary, scans it for new imperative or persona instructions, resumes from it, and compares behavior. Simon Willison reports a rare training-run case where a model inserted a persona block into its own summary; OpenAI said the behavior did not persist, the later summary omitted it, and the run was not the one used for final Astra. Simon Willison's Weblog

📡 WHAT SHIPPED

• Jev + LangChain integration: Jev is a non-generative “System One” model that returns typed answers and probabilities rather than text. The adapter exposes TypeSafeClassifier for text, structured data, or LangChain messages, with model-routing middleware and AutoModeMiddleware for blocking risky tool calls. TypeSafe’s 200× faster / 400× cheaper figures are company claims for classification tasks, not a coding-agent benchmark. @sydneyrunkle

• OpenWiki v0.5.2 adds IBM Bob Shell. Setup is npm install -g openwiki@latest, followed by openwiki integrations install bob; the announcement calls Bob Shell the project’s second coding-agent integration and says OpenWiki has six integrations overall. @colifran_

• Kody adds external secret providers. Kent C. Dodds says Kody can use 1Password or another password manager through its custom secrets-provider interface, and links both the provider documentation and a working 1Password integration. @kentcdodds

• LangSmith rebuilt agent-trace filtering. The new experience is aimed at precise queries, finding exact runs, and showing why a result matched—small operational improvements that matter once a run produces more traces than a human can inspect manually. @LangChain

🎬 GO DEEPER

• Theo — “Please stop using stupid models”: execution-based review and screenshot-to-PR. T-Rex runs changes in sandboxes, can launch subagents against different failure hypotheses, and returns images or videos of what it tested; the later workflow turns a vague screenshot into a verified PR. Theo - t3․gg

• Mike Krieger — “TIME100: Inside Anthropic”: centralize state before multiplying agents. Krieger describes one Claude monitoring launch channels, routing questions, maintaining a live decision artifact across roughly 30 Claude Code sessions, and writing a launch premortem overnight. Salesforce

• Repo to study — langchain-samples/deep-life-sci. Ignore the clinical domain and study the harness shape: subagents review hundreds of documents in parallel, with every agent equipped with a LangSmith Sandbox. @LangChain

Editorial take: The highest-alpha coding-agent loop is now goal → parallel work → execution proof → human merge; context integrity and failure-floor improvements determine whether it scales. @ClaudeDevs +2