​
Generic ACP brains (Roadmap)

The plan is to let a session run on an existing coding agent instead of the server harness, selected per agent and without changing the API your app uses:
  • Platform harness interface (phase H0). The server harness and the external-brain path are put behind one internal interface (submit input, cancel, close, capabilities). Behavior does not change. In progress.
  • ACP client in bmad (phase H1). bmad hosts the ACP agent process, translates ACP updates into Managed Agents events, and reports them through the agent capability. In progress.
  • First brain: DeepSeek Harness (DSH) via dsh --profile acp. A feasibility study ran it on Linux with real tool calls, approvals, cancellation, cross-process resume and concurrent sessions. Known limits to resolve: streaming arrives per committed message rather than per token, errors are unstructured text, usage reports only context occupancy, and compaction is not visible over ACP.
  • Next brains: Codex, then Claude Code. Running an ACP brain on BMA’s servers (rather than in a sandbox or on your computer) is not planned until a brain can delegate all of its tools to BMA’s hands.
On every brain, clients keep receiving the same Managed Agents events. A capability a brain lacks is reported explicitly, never faked.
Nothing on this page is available yet. Dates are not committed.