# Manager Agents, Event Triggers, and Claude Code's New Control Surface

*By Coding Agents Alpha Tracker • April 15, 2026*

Notion dropped the most transferable production patterns of the day, from progressive tool disclosure to manager agents and eval loops. Anthropic, Cursor, and LangChain added new ways to run coding agents in parallel, on events, and behind real auth boundaries.

## 🔥 TOP SIGNAL

Today's highest-alpha download came from Notion: the scaling limit was not finding a smarter model, it was stopping the habit of cramming more tools and few-shot examples into one giant agent prompt. Simon Last and Sarah Sachs describe 4-5 harness rebuilds since late 2022, then the shift to progressive tool disclosure, distributed tool ownership, manager agents, and evals-as-agent-loops once the system grew past 100 tools [^1].

## 🛠️ TOOLS & MODELS

- **Claude Code on desktop got a real control-surface upgrade.** Anthropic rebuilt it to run multiple Claude sessions side by side from one window with a new sidebar; Cat Wu says it is now the best way to manage local and cloud sessions, with git status, pinned sessions, and drag/drop layouts, and Alex Albert says Cowork + Code now cover most of his work [^2][^3][^4][^5].
- **Claude Code Routines pushes coding agents into automation.** You can trigger templated agents on a schedule, from GitHub events, or via API using Anthropic infra plus your MCP+repos; Anthropic says the feature already changed how it handles docs and backlog maintenance. Get started: [claude.ai/code/routines](http://claude.ai/code/routines) [^6][^7][^6].
- **Cursor Automations now hook into Sentry.** The workflow is straightforward: new issue arrives, the agent investigates root cause, opens a PR with a fix, and posts a Slack summary. Template: [cursor.com/marketplace/automations/investigate-sentry-issues](http://cursor.com/marketplace/automations/investigate-sentry-issues) [^8][^9].
- **DeepAgents keeps leaning into open agent infra.** `deepagents v0.5` / `deepagentsjs v1.9.0` adds async subagents that can run on any Agent Protocol server in parallel with the main agent, plus multimodal `read_file` support and better prompt caching for Claude models; `deepagents deploy` is positioned as an open alternative to Claude managed agents, with user memory and more subagent support coming soon [^10][^11][^12].
- **OpenClaw v2026.4.14 is a reliability release worth reading.** Highlights: smarter GPT-5.4 routing and recovery, Chrome/CDP improvements, stuck-subagent fixes, Slack/Telegram/Discord fixes, and performance work. Release notes: [github.com/openclaw/openclaw/releases/tag/v2026.4.14](https://github.com/openclaw/openclaw/releases/tag/v2026.4.14) [^13].
- **Practitioner model split, not consensus.** Kent C. Dodds says Claude Desktop currently beats ChatGPT for MCP, understanding, persistence through tool calls, and memory, and works better when he asks Kody to generate UI apps; Theo, by contrast, still likes Claude models for coding/UI quality but mostly uses GPT models for coding and dislikes Claude Code as a harness [^14][^15][^16].

## 💡 WORKFLOWS & TRICKS

- **If your agent is getting dumber as you add tools, stop showing it all the tools.** Notion hit the point where even greeting the agent cost thousands of tokens; the fix was progressive disclosure and tool search, with the team explicitly fighting to keep the prompt short even as the tool surface passed 100 [^17][^1].
- **Manager-agent pattern:** let specialist agents write issues/tasks to a shared database or invoke one another directly, then give one manager agent visibility across the fleet and route only aggregated blockers to humans. In Notion's example, that turned 70+ agent notifications per day into about 5 [^17][^1].
- **Run evals like a coding-agent job, not a spreadsheet ritual.** Simon Last's loop: agent downloads the dataset, runs the eval, iterates on failures, debugs, and implements the fix; Sarah Sachs says teams then keep those evals in CI or nightly runs so model or harness changes are visible fast [^17][^1].
- **Pick CLI or MCP based on what can go wrong.** Use CLI when you want self-debugging, bootstrapping, long-output navigation, and progressive disclosure inside the same terminal; use MCP when you want a narrower, lightweight agent with tighter permission boundaries [^1].
- **Move from prompting to triggering.** Anthropic's new Routines and Cursor's Sentry automation both push the same pattern: define a templated agent once, then launch it from schedules, GitHub events, APIs, or incidents instead of starting from a blank chat every time [^6][^7][^6][^8].
- **Production auth recipe for deployed agents:** in `langgraph.json`, point at your agent, `auth.py`, and routes; in `@auth.authenticate`, validate the token and return a minimal user object; in `@auth.on('resources.*')`, write `owner=user['id']` into metadata so threads auto-filter per user; gate `crons.create` by role; pass the access token in the `Authorization` header; test locally with `uv run langgraph dev`, then ship with `uv run langgraph deploy` [^18].
- **Quick Codex starter you can copy today:** open the native macOS app starter prompt, describe the app, and let Dimillian's plugin supply the UI defaults, run-button wiring, and telemetry. Starter: [developers.openai.com/codex/use-cases/native-macos-apps](https://developers.openai.com/codex/use-cases/native-macos-apps) [^19][^20].
- **Low-slop loop from ThePrimeagen:** codify your programming rules, use several stages, and keep the agent on small changes. He says the gain so far is modest speed, not magical productivity, which is exactly why the pattern feels trustworthy [^21].

## 👤 PEOPLE TO WATCH

- **Simon Last + Sarah Sachs:** best production-agent interview of the day. Why it matters: 4-5 harness rebuilds, 100+ tools, eval platform work, manager agents, and a grounded CLI-vs-MCP view from a team that has been iterating since 2022 [^1].
- **@_catwu:** high-signal Anthropic builder account right now. Today she surfaced both the new desktop workflow and Routines triggers, plus concrete internal use cases like docs and backlog maintenance [^4][^6].
- **Romain Huet:** useful if you want concrete Codex workflows instead of benchmark chatter; today's macOS app flow is immediately reproducible [^19][^20].
- **ThePrimeagen:** still one of the better anti-hype filters. His current stance is not AI-writes-everything but codify rules, stage the work, inspect the output, and change your mind slowly [^21][^22].
- **Theo:** worth tracking for harness-business-model reality. He says Anthropic explicitly forbids using OAuth-backed Claude Code subscriptions in third-party harnesses, while OpenAI and GitHub are more permissive [^16].

## 🎬 WATCH & LISTEN

- **Notion on tool design (49:39-51:40):** the best short explanation today of why model-facing interfaces should match what the model wants, not your internal data model. This is the JavaScript → XML → Markdown/SQLite evolution in one clip [^1].


[![Notion’s Sarah Sachs & Simon Last on Custom Agents, Evals, and the Future of Work](https://img.youtube.com/vi/ATt7QJgt-2k/hqdefault.jpg)](https://youtube.com/watch?v=ATt7QJgt-2k&t=2979)
*Notion’s Sarah Sachs & Simon Last on Custom Agents, Evals, and the Future of Work (49:39)*


- **Notion on manager agents (36:36-38:22):** concrete multi-agent ops: specialist agents file work and blockers, one manager agent watches the fleet, and the human only sees the compressed queue [^1].


[![Notion’s Sarah Sachs & Simon Last on Custom Agents, Evals, and the Future of Work](https://img.youtube.com/vi/ATt7QJgt-2k/hqdefault.jpg)](https://youtube.com/watch?v=ATt7QJgt-2k&t=2196)
*Notion’s Sarah Sachs & Simon Last on Custom Agents, Evals, and the Future of Work (36:36)*


- **LangSmith multi-tenant auth demo (09:25-10:48):** fast watch if you are deploying agents to real users; it shows local testing, user-scoped threads, and the final `langgraph deploy` step end to end [^18].


[![Make Your LangSmith Deployment Multi-Tenant](https://img.youtube.com/vi/DkNqgCz8cjE/hqdefault.jpg)](https://youtube.com/watch?v=DkNqgCz8cjE&t=565)
*Make Your LangSmith Deployment Multi-Tenant (9:25)*


## 📊 PROJECTS & REPOS

- **OpenClaw v2026.4.14:** open-source coding-agent stack shipping a reliability pass—smarter GPT-5.4 routing/recovery, better Chrome/CDP behavior, unstuck subagents, chat integration fixes, and perf work. Release notes: [github.com/openclaw/openclaw/releases/tag/v2026.4.14](https://github.com/openclaw/openclaw/releases/tag/v2026.4.14) [^13].
- **DeepAgents / deepagents deploy:** open deployment stack for long-running agents, with async subagents on Agent Protocol servers, multimodal file handling, and an explicit open-alternative-to-Claude-managed-agents pitch. Posts: [deepagents v0.5](https://blog.langchain.com/deep-agents-v0-5/) and [deploy](https://blog.langchain.com/deep-agents-deploy-an-open-alternative-to-claude-managed-agents/) [^10][^11].
- **Cursor's multi-agent CUDA-kernel project:** not a repo drop, but a real project signal. Cursor says the system achieved a 38% geomean speedup across 235 problems in 3 weeks, beat baselines on 63% of problems, delivered >2x speedups on 19%, and learned distinct optimization strategies on Blackwell 200 kernels from scratch. Research: [cursor.com/blog/multi-agent-kernels](http://cursor.com/blog/multi-agent-kernels) [^23][^24][^25][^26].

*Editorial take: the strongest signal today is that real coding-agent progress now looks like ops work—short prompts, event triggers, permission boundaries, async decomposition, and supervision layers around the model [^1][^6][^8][^18][^1].*

---

### Sources

[^1]: [Notion’s Sarah Sachs & Simon Last on Custom Agents, Evals, and the Future of Work](https://www.youtube.com/watch?v=ATt7QJgt-2k)
[^2]: [𝕏 post by @amorriscode](https://x.com/amorriscode/status/2044129923644961155)
[^3]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2044131493966909862)
[^4]: [𝕏 post by @_catwu](https://x.com/_catwu/status/2044212251717186007)
[^5]: [𝕏 post by @alexalbert__](https://x.com/alexalbert__/status/2044144091395699055)
[^6]: [𝕏 post by @noahzweben](https://x.com/noahzweben/status/2044093913376706655)
[^7]: [𝕏 post by @_catwu](https://x.com/_catwu/status/2044103103591657941)
[^8]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044097171071611338)
[^9]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044097172556394957)
[^10]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2044086454230626733)
[^11]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2044097913698091496)
[^12]: [𝕏 post by @sydneyrunkle](https://x.com/sydneyrunkle/status/2044097535929651209)
[^13]: [𝕏 post by @openclaw](https://x.com/openclaw/status/2044042546976883063)
[^14]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2044143935262732511)
[^15]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2044144112564351477)
[^16]: [Anthropic thinks they're Apple. They're actually hypocrites.](https://www.youtube.com/watch?v=ysS8GDwsaK8)
[^17]: [Notion’s Token Town: 5 Rebuilds, 100+ Tools, MCP vs CLIs and the Software Factory Future — Simon Last & Sarah Sachs of Notion](https://www.latent.space/p/notion)
[^18]: [Make Your LangSmith Deployment Multi-Tenant](https://www.youtube.com/watch?v=DkNqgCz8cjE)
[^19]: [𝕏 post by @romainhuet](https://x.com/romainhuet/status/2044207684111794399)
[^20]: [𝕏 post by @romainhuet](https://x.com/romainhuet/status/2044208029969924218)
[^21]: [𝕏 post by @ThePrimeagen](https://x.com/ThePrimeagen/status/2043861800819761382)
[^22]: [𝕏 post by @ThePrimeagen](https://x.com/ThePrimeagen/status/2044146948861554852)
[^23]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044136953239740909)
[^24]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044136955668230225)
[^25]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044136959061438889)
[^26]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044136960210677884)