# Background Loops, Agent Docs, and Multi-Model Routing

*By Coding Agents Alpha Tracker • May 5, 2026*

Today’s best signals are operational, not aspirational: top practitioners are turning coding agents into scheduled background workers. The practical edge is in loops, future-aware prompts, on-distribution stacks, repo-specific docs, and the infrastructure shipping around them.

## 🔥 TOP SIGNAL

Today's clearest signal: background agents are becoming the real coding workflow. Boris Cherny says he runs dozens of Claude Code `/loop`s to babysit PRs, fix flaky CI, and cluster feedback every 30 minutes, with new server-side `routines` keeping jobs alive when the laptop is closed [^1]. Alexander Embiricos says Codex already supports the same time-based pattern for unresolved discussions, launch bugs, and flaky tests [^2]—and Riley Brown's warning is the useful counter-signal: cronjobs, memory, re-auth, and file-placement reliability are still where agent power users lose time [^3].

## ⚡ TRY THIS

- **Steal Boris Cherny's first three loops.** Set up recurring jobs for `PR babysitting` (auto-rebase + fix CI), `CI health` (catch/fix flaky tests), and `feedback triage` (cluster feedback every 30 minutes). Run them on a cron-style repeat via `/loop`; if your tool has server-side execution, move long runners there so they keep working offline [^1].

- **Use future-oriented prompts as lightweight automation.** Embiricos says he uses this pattern in Codex all the time and that it's powerful but non-obvious [^2][^4].
  > "tomorrow, check in on this discussion and ping me if it isn't resolved" [^2]
  >
  > "let me know if this bug isn't fixed by the day before launch" [^2]
  >
  > "bug me if this flaky test doesn't go green after retry" [^2]

- **If you want max agent throughput, bias toward boring, on-distribution tech.** Cherny says Claude Code's codebase is simple TypeScript + React, originally chosen because that combo was very on distribution for the model; that helped them reach 100% model-written code early [^1]. If you're starting greenfield and expect heavy agent involvement, this is the pragmatic default [^1].

- **Write the migration doc before the port.** Bun appears to be exploring a Zig→Rust port with a dedicated [`docs/PORTING.md`](https://github.com/oven-sh/bun/blob/46d3bc29f270fa881dd5730ef1549e88407701a5/docs/PORTING.md) aimed at coding agents [^5][^6][^5]. Steal the pattern: if agents are handling a big refactor or language move, give them a repo-local playbook first [^5].

## 📡 WHAT SHIPPED

- **Bun showed two strong agent-native signals.** Armin Ronacher reported a bug and says a coding agent fixed it and pushed [PR #30257](https://github.com/oven-sh/bun/pull/30257) within five minutes; later, agents were debating on the PR itself [^7][^8]. Simon Willison separately spotted Bun's agent-specific [`PORTING.md`](https://github.com/oven-sh/bun/blob/46d3bc29f270fa881dd5730ef1549e88407701a5/docs/PORTING.md) as the project explores a Zig→Rust port [^5][^6][^5].

- **Vercel `deepsec`.** New open-source coding security harness: CLI-first, sandbox-based scaling, pluggable coding agents, large-repo focus, and support for AI Gateway or your own subscription. Vercel says it followed months of internal use and tests on some of the largest open-source codebases; blog: [introducing deepsec](https://vercel.com/blog/introducing-deepsec-find-and-fix-vulnerabilities-in-your-code-base) [^9].

- **`deepagents-cli` + Profiles API.** LangChain is pushing it as a model-agnostic harness for open-weight coding agents. Recent CLI features: `/agents`, `/model`, headless `--json + --max-turns`, `--acp`, `/skill:name`, and MCP with OAuth; docs: [overview](https://docs.langchain.com/oss/python/deepagents/cli/overview) [^10][^11][^10].

- **LangSmith Fleet multi-model routing.** Sub-agents can now use different models, with the stated goal of pushing simple work to fast/cheap models and saving stronger models for the hard parts; page: [Fleet](http://langchain.com/langsmith/fleet) [^12][^13][^12][^13].

- **Gemini API infrastructure updates.** Logan Kilpatrick says webhooks are live for long-running tasks including agents, and the Interactions API now returns more human- and agent-readable error messages [^14][^15].

- **Codex adoption looks real; Copilot economics look strained.** @linuz90 called Codex his favorite coding app and says it now handles 90%+ of his work despite earlier terminal/lock-in hesitation [^16][^17]. Theo says one Copilot message burned through 60M+ tokens/$30, and 15 messages totaled $221 of tokens under a flat-message plan he thinks GitHub cannot sustain [^18][^19].

- **Model preference is still split.** Cherny says Claude Code reached 100% agent-written code on a simple TypeScript/React codebase, with each Anthropic release from Opus 4 through 4.7 improving the curve [^1]. Theo, by contrast, says Anthropic is still meaningfully worse than OpenAI for most code outside frontend, even though many enterprise developers use Claude/Opus via Bedrock, Cursor, or Copilot in existing cloud setups [^20].

## 🎬 GO DEEPER

- **7:35-8:49 — Boris Cherny's `/loop` playbook.** Best short walkthrough today of a practical background-agent setup: PR babysitting, CI repair, feedback clustering, and why server-side `routines` matter once jobs need to survive laptop sleep [^1].
  
  
[![Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next](https://img.youtube.com/vi/SlGRN8jh2RI/hqdefault.jpg)](https://youtube.com/watch?v=SlGRN8jh2RI&t=455)
*Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next (7:35)*


- **19:50-20:37 — When the model starts the loop for you.** Cherny says Claude 4.7 increasingly notices time-varying work on its own, starts a loop, and offers 30-minute Slack reports [^1].
  
  
[![Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next](https://img.youtube.com/vi/SlGRN8jh2RI/hqdefault.jpg)](https://youtube.com/watch?v=SlGRN8jh2RI&t=1190)
*Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next (19:50)*


- **Study Bun's live artifacts, not the discourse.** [PR #30257](https://github.com/oven-sh/bun/pull/30257) is a report→fix→PR example that landed minutes after a bug report, and Bun's [`docs/PORTING.md`](https://github.com/oven-sh/bun/blob/46d3bc29f270fa881dd5730ef1549e88407701a5/docs/PORTING.md) shows what agent-facing migration guidance can look like in a real repo [^7][^8][^5].

- **Study Simon Willison's narrow-tool workflow.** His [Redis Array Playground](https://tools.simonwillison.net/redis-array) and [PR #277](https://github.com/simonw/tools/pull/277) show Claude Code for web being used for a focused dev utility around one new Redis feature, not a giant monolith ask. More context: [blog post](https://simonwillison.net/2026/May/4/redis-array/) [^21][^22][^21].

*Editorial take: the edge is shifting from single-shot codegen to reliable background workflows—loops, timers, sub-agent routing, and repo-specific guidance.* [^1][^2][^12][^5][^3]

---

### Sources

[^1]: [Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next](https://www.youtube.com/watch?v=SlGRN8jh2RI)
[^2]: [𝕏 post by @embirico](https://x.com/embirico/status/2051521360896422264)
[^3]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2051372403494949125)
[^4]: [𝕏 post by @embirico](https://x.com/embirico/status/2051521537543741494)
[^5]: [𝕏 post by @simonw](https://x.com/simonw/status/2051476878712840407)
[^6]: [𝕏 post by @LukeParkerDev](https://x.com/LukeParkerDev/status/2051431276205461635)
[^7]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2051410782546436427)
[^8]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2051416096289267862)
[^9]: [𝕏 post by @vercel_dev](https://x.com/vercel_dev/status/2051381241283539255)
[^10]: [𝕏 post by @masondrxy](https://x.com/masondrxy/status/2051359502918648319)
[^11]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2051360793904529439)
[^12]: [𝕏 post by @BraceSproul](https://x.com/BraceSproul/status/2051366289449570594)
[^13]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2051367244060598312)
[^14]: [𝕏 post by @OfficialLoganK](https://x.com/OfficialLoganK/status/2051434527931953490)
[^15]: [𝕏 post by @OfficialLoganK](https://x.com/OfficialLoganK/status/2051434948385742998)
[^16]: [𝕏 post by @linuz90](https://x.com/linuz90/status/2051273382327685207)
[^17]: [𝕏 post by @gdb](https://x.com/gdb/status/2051297950672859290)
[^18]: [𝕏 post by @theo](https://x.com/theo/status/2051218167780041147)
[^19]: [𝕏 post by @theo](https://x.com/theo/status/2051395816410210604)
[^20]: [Microsoft and OpenAI break up \(Amazon is pumped\)](https://www.youtube.com/watch?v=fxxpQhJyupQ)
[^21]: [𝕏 post by @simonw](https://x.com/simonw/status/2051348328991043863)
[^22]: [Redis Array Playground](https://simonwillison.net/2026/May/4/redis-array)