OpenHands went from agent to platform over agents in six months

OpenHands from agent to coordinator
https://www.openhands.dev/blog/introducing-agent-canvas
https://docs.openhands.dev/openhands/usage/agent-canvas/acp-agents
Previously OpenHands itself was an agent, now via Agent Client Protocol (ACP) it launches and serves harnesses like Claude Code, Codex and OpenHands itself on desktop. It was brought into the architecture back in spring, and over the summer it became the central scenario. Agent Canvas is now effectively the main interface from which you can run multiple agents in parallel, see files, commands, conversations, automations and attach remote backends. Added a clear context-window usage meter, a drawer with context usage info and manual compaction.

A new Agent Profile object appeared — a saved configuration of how exactly to start an agent. It holds agent type, model, credentials and scoped secrets, MCP servers, skills and more. Automations turned from a simple feature into an orchestrator with Agent Profile selection, export/import as JSON, configurable timeouts, conversation launch via plugin, cron, live execution phase, task result understanding, tags, view/manage permissions, auto-disable after a series of errors. Moving toward agent errors being caught automatically.

#openhands #acp #selfhost