# Durable Agent Workflows, Codex+Notion Playbooks, and New LangChain Infra *By Coding Agents Alpha Tracker • June 2, 2026* Today's strongest signal is architectural: coding agents are becoming long-running workflows with real state, sleep, review, and safe execution primitives. Also covered: Riley Brown's copyable Codex+Notion patterns and the latest releases from LangChain, Google, and Cursor. ## 🔥 TOP SIGNAL Production coding agents are crossing a line from chat-with-tools to long-running workflows with real runtime design. Addy Osmani distilled the core requirements into three parts: true dormancy, durable checkpoints on every transition, and a separate evaluator instead of letting the agent grade its own work [^1]. LangChain's newest agent infra points the same direction—managed cross-session context plus isolated sandboxes with persistence, auth controls, and snapshot/restore [^2][^3]. ## ⚡ TRY THIS - **Move instructions out of chat and into durable docs.** Riley Brown's setup: install the Notion plugin in Codex, open Notion inside Codex's signed-in browser, put a top-of-page banner like `if you are an AI agent, read the following tabs`, and use `Cmd+Cmd` App Shots to pass the exact page context before asking for edits [^4]. Kent C. Dodds' lighter variant is simpler: keep markdown files around and let the agent find the right instructions on demand [^5]. - **Only promote repeat work to a skill after you've seen a perfect run.** Riley's method: make Codex do the task, tighten formatting, links, and conciseness until it behaves, then say `make that a skill` or `turn that into a skill called Notion Quick Note` [^4]. He uses the same pattern for `Notion Research`, where Slash Tabs add research at the top of a page without cluttering the main document [^4]. - **Give the agent its own notebook and a nightly recap.** Riley keeps a separate Notion notebook that Codex can write to, then asks it to create a 10pm automation to write a one-page daily summary, infer the top tasks, and email it [^4]. He says this became useful once he started generating many long chat threads per day [^4]. - **For long-running jobs, wire explicit wake-up and review paths.** Peter Steinberger tells Codex to call [sag.sh](http://sag.sh) whenever it needs human help—for example, a release blocked on 1Password—so the agent asks only when it is stuck [^6]. Pair that with Addy Osmani's production pattern: sleep until a webhook, schedule, human callback, or tool callback wakes the agent, checkpoint state on every transition, and split planner, generator, and evaluator roles so the writer is not its own reviewer; his caution is that agents still need human judgment for the final 20-30% [^1]. ## 📡 WHAT SHIPPED - **Google / Addy Osmani:** Addy said his team shipped ADK 2.0 plus a graph-based Agent CLI runtime, prepackaged skills, Gemini 3.5 Flash, and AntiGravity 2.0; he also pointed to open-source long-running-agent docs and new Skills Registry docs [^1]. - **LangChain — Managed Deep Agents:** keeps the familiar `AGENTS.md`, `skills/`, `subagents/`, and `tools.json` shape, while `Context Hub` stores and updates context across sessions so agent definitions can evolve over time. Blog: [Managed Deep Agents](https://www.langchain.com/blog/introducing-managed-deep-agents) [^2][^7]. - **LangSmith Engine:** LangChain is pitching this as a way to stop manual failure triage: connect your tracing project, optionally connect your repo, then review and merge suggested improvements. Link: [LangSmith Engine](https://www.langchain.com/langsmith/engine) [^8][^9]. - **LangSmith Sandboxes:** LangChain's keynote framed this as safe execution infra with isolated runtime, network controls, persistent state, and snapshot/restore [^10]. Mukil Loganathan added the concrete product details: about 0.98s P50 spin-up, dynamic scale to thousands, an auth proxy with allow and deny lists plus credentials kept out of the runtime, pause and resume, no lifetime limit, multi-agent shared state, bring-your-own Docker or CLI support, and paid-plans-only availability; roadmap items include local and remote handoff, shared volumes, and full execution tracing [^3]. - **LangSmith LLM Gateway:** LangChain posted a 3-step setup for routing and policy control—point agents at the gateway with a LangSmith API key, add provider keys to workspace secrets, then set policies in the UI. Blog: [LLM Gateway](https://www.langchain.com/blog/introducing-llm-gateway) [^11][^12]. - **Cursor Teams:** usage limits are going up for every Teams user, and a new Premium team seat offers 5x usage for 3x the cost. Announcement: [teams pricing update](https://cursor.com/blog/teams-pricing-june-2026) [^13][^14]. - **Adoption signal:** Mukil estimated that roughly 70% of the Interrupt audience already uses coding agents; he also said LangChain's internal OpenSUI commits hundreds of PRs across repos, while citing Google at 75% AI-generated code and Stripe at 1,300 AI-generated PRs per week [^3]. ## 🎬 GO DEEPER - **1:47-4:33 — Addy Osmani on long-running agent architecture.** Covers sleep, checkpoints, and evaluator separation in one sequence [^1]. - **9:29-12:13 — Riley Brown on skill bootstrapping in Codex.** A clean walkthrough of the `do it well once -> make that a skill` loop, using `Notion Quick Note` as the example [^4]. - **6:31-8:58 — Mukil Loganathan on sandbox safety primitives.** Covers auth proxy, persistent state, pause and resume, and snapshot/restore for untrusted agent code [^3]. - **Project and doc pages worth reading:** [Managed Deep Agents](https://www.langchain.com/blog/introducing-managed-deep-agents) for cross-session context [^7], [LangSmith Engine](https://www.langchain.com/langsmith/engine) for failure-repair workflow [^9], and [LLM Gateway](https://www.langchain.com/blog/introducing-llm-gateway) for policy and routing setup [^12]. [![Build long-running agents with Google’s Agentic Stack | The Agent Factory](https://img.youtube.com/vi/VGpfnbr7rso/hqdefault.jpg)](https://youtube.com/watch?v=VGpfnbr7rso&t=139) *Build long-running agents with Google’s Agentic Stack | The Agent Factory (2:19)* [![Learn To Use Notion With AI Agents (Full Guide)](https://img.youtube.com/vi/YGWcFMR8gk8/hqdefault.jpg)](https://youtube.com/watch?v=YGWcFMR8gk8&t=569) *Learn To Use Notion With AI Agents (Full Guide) (9:29)* *Editorial take: the frontier is shifting from clever one-shot prompts to agent runtime design—persistent instructions, durable state, human callbacks, and separate evaluators are what turn agent demos into dependable workflows.* [^4][^1] --- ### Sources [^1]: [Build long-running agents with Google’s Agentic Stack | The Agent Factory](https://www.youtube.com/watch?v=VGpfnbr7rso) [^2]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2061432934993674267) [^3]: [Run Untrusted Agent Code with LangSmith Sandboxes | Interrupt 26](https://www.youtube.com/watch?v=IIchUA5T3gs) [^4]: [Learn To Use Notion With AI Agents \(Full Guide\)](https://www.youtube.com/watch?v=YGWcFMR8gk8) [^5]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2061494089309393379) [^6]: [𝕏 post by @steipete](https://x.com/steipete/status/2061574752574283858) [^7]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2061510459082121443) [^8]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2061492038722294027) [^9]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2061492307283558667) [^10]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2061448130806116827) [^11]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2061463141494489101) [^12]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2061463142333354041) [^13]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2061550723503194426) [^14]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2061550724736241718)