How agents check that the OpenHands Agent SDK does what it says, one feature at a time: a map of every behavior an agent can see, a CLI that drives the real thing, and known bugs kept as expected failures until someone fixes them. Everything here is open, and anyone can rerun it.
The explainer
Two minutes on what an agent actually sees when it uses the SDK: the map, the control-agent-server CLI, a recipe that caught a bug, and the full replay. Drawn with Manim, with a generated score. Download (MP4, 10 MB) · Scene script and music
Agent Server
The first map: 34 families and 931 sub-features of the Agent Server, driven over REST, WebSocket, the Python client and the TypeScript client. Its full replay: 771 pass, 156 known bugs reproduced, 6 blocked, no surprises. It lives in software-agent-sdk PR #5621.
The next map covers the SDK used in process: the agent loop, LLMs, conversations, condensers, skills, MCP, tools, workspaces and more, in 21 families (S01–S21). It is being drafted now and will be linked here when it is published.
OpenHands app
The sister project: the same approach applied to the OpenHands app (Agent Canvas), with a browser-driving CLI and real before/after screenshots.