Started April 9. As of today: 150k LOC TypeScript, 70 merged PRs, 147 test files. It's a wire-compatible open-source implementation of Anthropic's Managed Agents API ( repo , Apache 2.0) running on Cloudflare Workers + Durable Objects + Containers. Solo, basically no hand-written code.

Just Claude Code. No Cursor, no Aider, no orchestration framework.

Console UI is React + Tailwind + shadcn, also vibe-coded. The design quality is from impeccable.style skills loaded into Claude Code. Without them the UI looks like every other shadcn dump. With them:

https://preview.redd.it/lgkqr8qyyd3h1.png?width=4808&format=png&auto=webp&s=fae4546b9818b31e50396acd58e41cdebfaebf24

Four things that made this work for me:

1. Spec-driven. Anthropic published the Managed Agents API spec with every endpoint and event type spelled out. I didn't have to invent the surface area; just implement it. The agent always has a concrete "what should this do" to look up. Honestly this is why it scaled at all. Vibe coding falls apart fastest when "what to build" is ambiguous — if you're picking what to vibe-code next, pick something with a published spec.

2. Test-driven, with the agent writing the tests. 147 test files, all agent-generated. The loop: agent writes a failing integration test against the spec → I review the test → agent implements → CI runs it → green. Tests end up being how I talk to the agent, not a chore. They catch the AI's confidently-wrong moments before they ship.

3. API + CLI first, designed for an agent to drive. openma itself is API-first: oma agents create / sessions chat / memory write / bridge setup / ... ( npm ). CLI was built for an agent to operate openma, not for humans to click through a dashboard. Same reason the rest of the stack worked: wrangler (Cloudflare) and gh (GitHub) are both agent-friendly CLIs, so the agent deploys, tails logs, manages PRs and releases without me. I basically never leave the terminal.

4. Parallel CI lanes on Cloudflare + GitHub. Each feature gets a git worktree, a branch, and a per-PR Cloudflare preview deploy. The full integration suite runs against actual edge infrastructure (real DOs, R2, Queues, not simulators) on every PR. So I can have 5-10 features in flight at once without anyone blocking anyone else, and each lane gets validated against real CF behavior before merge.

One thing that hasn't worked:

Engineering taste — the agent has none. In v1, every piece of state ended up in Cloudflare KV: agent configs, sessions, event logs, even uploaded files. KV is just the easiest API to reach for, so the agent kept stuffing more into it. Took an explicit architecture review to split things back out properly (KV → config only, D1 → indexed/transactional state, R2 → blobs, DO SQLite → event log). Don't trust the agent on tech selection. Review the architecture early, and keep reviewing it. The wrong primitive compounds for weeks before you notice.

What's your stack at this scale? Especially curious how others structure parall…

为什么值得关注

能改变理解方式，而不只是重复常识；有直接可用的方法、工具或操作价值；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：tech，保留分：0.70