# Claude Tag Moves Coding Agents Into Slack; LangSmith Engine Closes the Repair Loop *By Coding Agents Alpha Tracker β€’ June 24, 2026* Anthropic's Claude Tag is the strongest workflow signal today: multiple insiders describe a Slack-native coding agent that lives in threads, handles incidents, and writes a large share of code. Also worth attention: LangSmith Engine's trace-to-PR loop, Cursor's new team skill surfaces, and a few OSS agent-infra projects getting real discussion. ## πŸ”₯ TOP SIGNAL Anthropic's **Claude Tag** is the clearest workflow shift today: Boris Cherny says the Slack-native agent launched today, and internally it's already used to write PRs, address user feedback, investigate incidents, run data analyses, and answer company-knowledge questions; he says **65% of the product team's new code** is created by Anthropic's internal version [^1][^2][^3]. Anthropic engineer @_catwu says the internal system **merges 65% of product PRs**, and Andrej Karpathy's reaction is that this is not Slack-wrapped RAG but a **deeply integrated, multiplayer** coding-agent product that changes how teams work [^4][^5]. > "it’s not some LLM Q&A with RAG over Slack... it’s a different way of working entirely... I work from Slack now." [^5] ## ⚑ TRY THIS - **Run incident response in the same thread the humans use.** @_catwu's flow: when the page lands, tag Claude in the incident thread; it pulls graphs, diffs the deploy, identifies root cause, and tags the author. The team approves in-thread, then Claude opens the fix, lands it, watches the metric recover, and resolves the page [^6]. - **Use a per-thread agent as both search and executor.** Boris Cherny's setup starts with `@.Claude` in a Slack channel. Each thread gets its own sandbox, memory, and permissions; Claude can clone repos, write/test/compile code, answer questions like `What’s the status on X?` or `Who owns this service?`, proactively watch channels, draft PRs, and react with `βœ…` / `❌` when a thread is resolved [^7][^8][^2][^7]. - **Steal LangChain's closed-loop improvement pattern.** The durable sequence is simple: **(1)** run the agent and mine weaknesses, **(2)** propose harness improvements, **(3)** confirm the changes help without regressions, then loop [^9]. LangSmith Engine turns that into an operational workflow: build the agent, test with datasets/evals, deploy + monitor traces/online evals, then patch, regress-test, redeploy; when traces show errors, eval failures, negative feedback, or new bad behaviors, Engine clusters them into issues and drafts targeted PRs plus new tests/evals for review [^10]. - **For messy browser/API exploration, have the agent build the harness first.** Simon Willison used Claude Code for web to build a small playground UI for testing OPFS + Pyodide behavior across browsersβ€”a good pattern when you need a disposable test bench before committing to product code [^11]. ## πŸ“‘ WHAT SHIPPED - **Claude Tag (Slack launch)** β€” Launched today; tag Claude in-channel and each thread gets its own sandbox, memory, and permissions. Cherny says it can proactively monitor channels, and Anthropic built security layers around model training, classifiers/auto mode, access controls, and channel/workspace boundaries. More: [introducing-claude-tag](https://www.anthropic.com/news/introducing-claude-tag) [^1][^8][^7][^12][^13] - **LangSmith Engine (available now)** β€” Connect a tracing project and Engine will inspect traces, cluster repeated failures into issues, draft targeted PRs, and propose new examples/evals for the test suite. LangChain says teams like Vanta, Campfire, and Cogent are already catching regressions earlier and cutting triage time [^10]. - **Cursor customize updates** β€” Plugins can now ship prebuilt canvases (example: Atlassian canvas for live issues/projects/docs); Cursor also added a team leaderboard for popular plugins/skills/MCPs with one-click add, and extended team marketplaces to GitLab, Bitbucket, and Azure DevOps alongside local repos. Changelog: [cursor.com/changelog/customize](http://cursor.com/changelog/customize) [^14][^15][^16] - **Google Interactions API GA** β€” One API for Gemini models and agents, with dedicated coding-agent skills, the Antigravity Agent remote Linux sandbox, multimodal tool use, and `background=True` for async long-running interactions [^17]. - **Notable comparison: GLM 5.2 vs Opus 4.8** β€” Jason Zhou shared a same-prompt, same-reference-image frontend test where both models produced a working Three.js logistics dashboard. He also cites GLM 5.2 pricing at **$1.40 / 1M input** and **$4.40 / 1M output**, with Opus about **5x more expensive** [^18][^19]. - **OSS agent infra worth a look** β€” In Matthew Berman's roundup, **Deer Flow** (~74k GitHub stars) stood out as a long-horizon harness built around sub-agents, memory, sandboxes, and skills; **Codebase Memory MCP** (~12k) claims Linux-kernel-scale indexing in 3 minutes with sub-millisecond structural queries and 120x fewer tokens; **Skill Specter** (<10k) scans skills for 65 vulnerability patterns before install [^20]. ## 🎬 GO DEEPER - **1:36–2:19 β€” LangSmith Engine's trace β†’ issue β†’ PR loop.** Best short walkthrough today of a real agent-improvement pipeline: watch traces, cluster patterns, draft the fix, add evals, then keep monitoring after merge [^10]. [![The Agent for Your Agent.](https://img.youtube.com/vi/VKFKyrrK-Iw/hqdefault.jpg)](https://youtube.com/watch?v=VKFKyrrK-Iw&t=95) *The Agent for Your Agent. (1:35)* - **6:20–7:26 β€” Codebase Memory MCP on huge repos.** If your agents keep burning tokens just to understand code structure, this is the pitch to examine: Linux kernel indexed in 3 minutes, structural queries in under 1 ms, 158 languages, 11 harnesses [^20]. [![You NEED to try these 12 open-source AI projects RIGHT NOW](https://img.youtube.com/vi/2lmBj_XQq0I/hqdefault.jpg)](https://youtube.com/watch?v=2lmBj_XQq0I&t=379) *You NEED to try these 12 open-source AI projects RIGHT NOW (6:19)* - **Study [Deep Agents](https://github.com/langchain-ai/deepagents) alongside LangChain's [loop engineering write-up](https://www.langchain.com/blog/the-art-of-loop-engineering).** The repo matters, but the bigger takeaway is the pattern: weakness mining, harness edits, regression checks, repeat [^9][^21]. *Editorial take: the serious setups are starting to look the sameβ€”put the agent inside the real workstream, give it bounded tools and memory, then close the loop with traces, tests, and human approval.* [^8][^6][^10] --- ### Sources [^1]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2069474681749754272) [^2]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2069474687323893796) [^3]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2069474683372839253) [^4]: [𝕏 post by @_catwu](https://x.com/_catwu/status/2069473118742331608) [^5]: [𝕏 post by @karpathy](https://x.com/karpathy/status/2069601818540392669) [^6]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2069468902216945939) [^7]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2069474688619958517) [^8]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2069474689819480394) [^9]: [𝕏 post by @sydneyrunkle](https://x.com/sydneyrunkle/status/2069476285374464380) [^10]: [The Agent for Your Agent.](https://www.youtube.com/watch?v=VKFKyrrK-Iw) [^11]: [OPFS + Pyodide test harness](https://simonwillison.net/2026/Jun/23/opfs-pyodide) [^12]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2069474685948186797) [^13]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2069474691010707486) [^14]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2069512595766173857) [^15]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2069512593887092811) [^16]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2069512597628440908) [^17]: [𝕏 post by @_philschmid](https://x.com/_philschmid/status/2069108134044467487) [^18]: [𝕏 post by @aibuilderclub_](https://x.com/aibuilderclub_/status/2068697779883606492) [^19]: [𝕏 post by @jasonzhou1993](https://x.com/jasonzhou1993/status/2069386776997990699) [^20]: [You NEED to try these 12 open-source AI projects RIGHT NOW](https://www.youtube.com/watch?v=2lmBj_XQq0I) [^21]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2069476688262520913)