Amazon's agent swarm rebuilt 200+ AWS services on a laptop in 4 nights
blaizedsouza · x · 2026-11-14
- Amazon shared the internal AI agent swarm setup it uses to build a working local copy of 200+ AWS services (438 public API models) entirely autonomously — no human wrote or reviewed the code, released on GitHub under CC0.
How it works
- The spec is AWS's public API descriptions; the reference is the real AWS cloud, used to validate every test
- Team 1 (test writers): builds tens of thousands of fidelity tests, each verified against real AWS
- Team 2 (issue finders): replays the tests against the emulator and files an issue per failure, grouping failures by root cause
- Team 3 (fixers): one agent per issue, finds the root cause and fixes it
- Merge gate: every fix must compile, pass coding-standard and security review agents, and break no tests — otherwise it's sent back
Hundreds of agents fix thousands of issues per day this way. The starting prompt was minimal; handed to Kiro, the first Amazon SQS came back in 15 minutes and the rest of AWS finished over a weekend.
The lesson: rebuilding the replication engine by hand took Swami Sivasubramanian and a colleague 4 months; the agents flailed for 4 nights without a way to check their work, but once given a spec and a test environment, it took 2 hours. The approach applies to any project with a spec and something real to check against: legacy rewrites, code ports, API cloning.
More from coding & agent
- Tencent open-sources LoopForge to self-host coding-agent workflows with approval gates — Tencent · 2026-11-14
- AG-UI, the agent-user interaction protocol for frontend apps, hits 17k GitHub stars — ag-ui-protocol · 2026-11-14
- The Test Engineer prompt: have AI design full unit, integration, and E2E test strategies — hey_abusiddik · 2026-11-14
- 8 copy-paste prompts for using Claude like a senior engineer, not an intern — hey_abusiddik · 2026-11-14
- Gemma 4 31B leaks Korean/Arabic/Cyrillic into Chinese-Vietnamese translation; fixes fail — LambdaHominem · 2026-11-14
- Muse Spark 1.3 has six reasoning levels — are you budgeting them per agent step? — Suresh935 · 2026-11-14