# Structured Agent Workflows Meet Production Reality *By Coding Agents Alpha Tracker β€’ June 6, 2026* This brief covers Addy Osmani's SDLC-shaped coding-agent workflow, MongoDB's rollout metrics across ~2,000 engineers, Simon Willison's sandbox build/test pattern, and the most relevant new releases around design-mode UIs, streaming, skills, and deployment. ## πŸ”₯ TOP SIGNAL Serious coding-agent users are converging on *structured* workflows, not blind autonomy. Addy Osmani's open-source skills system maps directly onto `define β†’ plan β†’ build β†’ verify β†’ review β†’ ship`, with each skill carrying usage guidance, red flags, and verification steps; Simon Willison and swyx describe the same pattern in practice by tightening review, approvals, and testing as stakes go up [^1][^2][^3]. MongoDB's rollout makes the stakes concrete: across ~2,000 engineers, coding assistants wrote 70% of last week's checked-in code and engineers produced ~35% more code, but token spend rose enough that the company is explicitly weighing tool cost against future hiring plans [^4]. ## ⚑ TRY THIS - **Addy Osmani β€” run a six-step loop before you let the agent code.** In VS Code/Copilot, start with `/refine` on a vague idea; answer the clarifying questions; run `/spec` to turn that into success criteria and stack assumptions; run `/plan` to generate phased tasks with acceptance criteria and suggested files so the agent stays inside a smaller blast radius; then do `/build`, `/verify`, and `/review` before shipping [^1]. Addy's build phase encodes TDD explicitly, and verify/review cover browser testing, security, simplification, CI, git, and docs [^1]. - **Addy Osmani + Command Code β€” keep context thin and route models by phase.** Put reusable behavior in small skill or taste markdown files. Let the agent see only the skill name/description first, then load full content on demand; Addy prefers this progressive loading for context-window economy and reaches for MCP more in auth-heavy cases [^1]. Then split models by job: Addy plans with Gemini 3.5 Flash and implements with stronger models like Gemini Pro, Opus, or Codex, while Ahmad Awais says users seed a project's in-repo taste file with Opus or GPT-5.5 and continue iterating with cheaper models [^1][^5]. - **Simon Willison β€” for high-stakes code, split research, build, and red-team passes.** Simon first had GPT-5.5 Pro research MicroPython WASI support and produce `research.md`, then handed that doc to Codex Desktop + GPT-5.5 high to build the package [^6][^7]. For the sandbox itself, he watched the trace closely and then used GPT-5.5 to try to break out of it, fixing edge cases it found [^2][^7]. > `read the research.md document and build this. You will probably need to write a script that compiles a custom WASM version of MicroPython as part of this project - fetch the MicroPython code to a /tmp directory for this as part of that script.` [^7] - **swyx β€” ask for pushback first, then keep autonomy bounded.** A simple prompt trick: frame the task as a questionβ€”literally append `?`β€”so the model critiques the idea and suggests alternatives instead of blindly executing [^8]. In his own conference-scheduling agent, he paired that with a bounded while-loop, minimal turns, human approval for proposals, and logs/confirmations across Slack, email, and web [^3]. That guardrail matters because, as Kent C. Dodds notes, agents can make you more confident you're on the right path when you're not [^9]. ## πŸ“‘ WHAT SHIPPED - **Simon Willison β€” `micropython-wasm` alpha + `datasette-agent-micropython` 0.1a0.** Safe Python execution in a MicroPython + WASM sandbox via wasmtime; Simon says he is using it himself after breakout testing with GPT-5.5 xhigh; repos: [micropython-wasm](https://github.com/simonw/micropython-wasm), [datasette-agent-micropython](https://github.com/datasette/datasette-agent-micropython) [^7]. - **Cursor β€” Design Mode.** Point, draw, or talk to update UI; blog: [cursor.com/blog/design-mode](http://cursor.com/blog/design-mode) [^10][^11]. - **LangChain β€” Managed Deep Agents.** Managed, model-agnostic infra for deploying deep agents with one line of code; details: [introducing-managed-deep-agents](https://www.langchain.com/blog/introducing-managed-deep-agents) [^12]. - **Deep Agents v0.6 β€” Streaming.** Adds a subscription model for tool and subagent progress in highly parallel systems; examples: [streaming-cookbook](https://github.com/langchain-ai/streaming-cookbook) [^13]. - **Skills are productizing fast.** LangSmith Fleet lets domain experts create reusable skills that stay in sync across teams, while Codex now reads `.agents/skills` and plans to deprecate `.codex/skills` [^14][^15]. - **Google ADK β†’ LangSmith deployments.** A wrapper turns an ADK runner into a deployment with persistence, streaming, tracing, and built-in endpoints; the demo flow is `wrap(...)` β†’ `uv run langgraph dev` β†’ `uv run langgraph deploy` [^16]. - **Command Code (emerging project).** Ahmad Awais says its per-repo `taste` files learn micro-preferences from accepts, edits, and rejects on merge, while its deterministic tool-repair layer has expanded to 16k+ variations across ~600B tokens; one user reportedly ran 12+ hour DeepSeek sessions and 70B tokens through it, and open source is planned very soon [^5]. ## 🎬 GO DEEPER - **11:21-15:29 β€” Addy Osmani: vague idea β†’ clarifying questions β†’ spec.** Watch `/refine` turn a rough GitHub-inspired habit tracker into concrete constraints and candidate directions before any real build work starts [^1]. [![Agent Skills | LIVE150](https://img.youtube.com/vi/Q6LR0mXArIQ/hqdefault.jpg)](https://youtube.com/watch?v=Q6LR0mXArIQ&t=681) *Agent Skills | LIVE150 (11:21)* - **18:16-21:24 β€” MongoDB: real rollout numbers from ~2,000 engineers.** CJ walks through 70% AI-written check-ins, ~35% more code, and the cost calculus that follows when token spend starts climbing [^4]. [![Agents in the Enterprise with MongoDB | Interrupt 26](https://img.youtube.com/vi/k4l-rtwezVg/hqdefault.jpg)](https://youtube.com/watch?v=k4l-rtwezVg&t=1095) *Agents in the Enterprise with MongoDB | Interrupt 26 (18:15)* - **7:40-8:16 β€” Simon Willison: when to trust the agent vs watch it like a hawk.** Short, useful distinction between low-stakes prototyping and security-critical code where you should scrutinize every step and then run adversarial tests [^2]. [![Scott and Mark learn to Vibe Check with Simon Willison | LIVE113](https://img.youtube.com/vi/-b1e0tyMVvY/hqdefault.jpg)](https://youtube.com/watch?v=-b1e0tyMVvY&t=460) *Scott and Mark learn to Vibe Check with Simon Willison | LIVE113 (7:40)* - **Repo to study β€” [micropython-wasm](https://github.com/simonw/micropython-wasm), [datasette-agent-micropython](https://github.com/datasette/datasette-agent-micropython), and Simon's [research.md](https://github.com/simonw/micropython-wasm/blob/c08fbd2276b15dc8c9bdff82845f750971f45647/research.md).** Good reference for research-model-first, coding-agent-second workflows plus WASM sandbox design with host functions and fuel limits [^7][^6]. - **Repo to study β€” [streaming-cookbook](https://github.com/langchain-ai/streaming-cookbook).** If you're building multi-tool or multi-subagent systems, this is the fastest way to see what useful streaming telemetry should look like in practice [^13]. *Editorial take: the biggest edge right now is not 'more autonomous agents'; it's tighter scaffolding around themβ€”phases, slim context, approvals, and hostile verification.* --- ### Sources [^1]: [Agent Skills | LIVE150](https://www.youtube.com/watch?v=Q6LR0mXArIQ) [^2]: [Scott and Mark learn to Vibe Check with Simon Willison | LIVE113](https://www.youtube.com/watch?v=-b1e0tyMVvY) [^3]: [Scott and Mark learn to Vibe Check with Swyx | LIVE115](https://www.youtube.com/watch?v=K893N9c0upY) [^4]: [Agents in the Enterprise with MongoDB | Interrupt 26](https://www.youtube.com/watch?v=k4l-rtwezVg) [^5]: [⚑️Making DeepSeek v4 outperform Opus 4.7 with Taste β€” @AhmadAwais , CommandCode.ai](https://www.youtube.com/watch?v=-rIAVuaRjOg) [^6]: [Running Python code in a sandbox with MicroPython and WASM](https://simonwillison.net/2026/Jun/6/micropython-in-a-sandbox) [^7]: [Running Python code in a sandbox with MicroPython and WASM](https://simonw.substack.com/p/running-python-code-in-a-sandbox) [^8]: [𝕏 post by @swyx](https://x.com/swyx/status/2063082950317486133) [^9]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2063011427808694285) [^10]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2062950344687272144) [^11]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2062950345928749525) [^12]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2062867544365560178) [^13]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2062942948518490405) [^14]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2062927890262483177) [^15]: [𝕏 post by @embirico](https://x.com/embirico/status/2018415923930206718) [^16]: [Deploy Google ADK Agents on LangSmith Deployment](https://www.youtube.com/watch?v=vxxxkNF5SXk)