Agentic-Flow ships the AgentDB dependency surface its CLI actually needs

What Agentic-Flow is

Agent orchestration with routing that learns: 66 agents, a large MCP tool surface, and hooks that decide where a task should go instead of sending everything to the largest model.

What changed

Pull request #181 ships the AgentDB CLI dependency surface, which is what makes the memory commands work from a clean install rather than only in the development checkout. It lands with the 2.1.1 release and follows governed MetaHarness promotion plus MCP profiles.

Get started

Node.js 18 or newer.

npx agentic-flow init
npx agentic-flow hooks pretrain

Expected result: an initialised project and a pretrained routing table.

Route a task and inspect what the router learned:

npx agentic-flow hooks route "implement user authentication"
npx agentic-flow hooks metrics

For an MCP host:

npx agentic-flow mcp start

Use it today

Practical case: cutting model spend on a busy repository. Input is your normal task stream; the workflow is pretrain, then let the router choose; the output is cheaper models handling the tasks that never needed the expensive one.

Acceptance test: record metrics for a week with routing on and a week with everything pinned to one model. Compare cost and rework. If routing does not win on at least one axis, turn it off.

Push it further

Experimental commentary. Dispatching a worker to deeply learn a topic and keep the result is the ambitious use. The limitation is evaluation: a learned route is only as good as the signal you feed back into it.

Falsifiable test: corrupt the feedback signal deliberately for a day and see whether routing quality degrades. If nothing changes, the loop is not actually closed.