AgentPlayground
Docs
The coordinator

The coordinator (its internal name is the Playground Keeper) is the single assistant you're
meant to talk to. It's the "central intelligence" of your install — everything flows through it,
and it has up to 25 reasoning steps to work through any one request before it has to stop and report
back.

What it's responsible for
  • Understanding intent — is this a quick one-off task, a recurring job, or a lasting project?
  • Delegating instantly — for a single well-defined task, it calls a team directly and returns that team's result. The team runs its own tool loop (scoped to that team's own tools) rather than the coordinator doing the work itself.
  • Planning across teams — for a multi-step goal, it can draft a plan, have it reviewed, then dispatch tasks to several teams in parallel and synthesize their outputs into one answer.
  • Managing the system — creating teams, agents, and skills when you ask for new ones.
  • Reading and writing the Brain — it searches the vault before answering domain questions, and writes results back so future tasks have more context.
  • Suggesting next steps — after finishing something, it's meant to propose what to do next rather than just stopping.
How delegation actually works

When you describe a task, the coordinator picks one of two paths:

  1. Direct delegation (delegate_to_team) — for a single clear task. It hands the target team a title and a description; that team executes its own tool loop and the coordinator reports back what was actually produced, not just that it "delegated."
  2. A reviewed plan (create_planrun_plan) — for anything that spans multiple steps or teams. A plan is drafted, a review pass runs over it, and once approved it dispatches tasks to every team in parallel and pulls the results together.

The coordinator can also delegate to more than one team in the same turn — for example fanning a
goal out across two teams at once rather than running them one after another.

The Filekeeper

One seeded team, Filekeeper, is specifically for Brain and file hygiene: sorting notes into the
right folders, fixing stray or duplicate files, and maintaining a running one-page summary of what
every team and project is currently doing. If you ask the coordinator to organize, tidy up, or keep
track of what's in progress, it delegates to Filekeeper rather than trying to do file cleanup
itself. Filekeeper never deletes a note without asking you first.

What it can reach
  • Delegation & plans — delegate to a team, create/run a plan, read a task or plan result.
  • The Brain / vault — semantic search, read a note, write or update a note.
  • The web — search and read pages.
  • System management — create teams, agents, and skills; schedule tasks and meetings; create and track projects.
  • Business skills — invoice generation, CRM contact tracking, proposal writing, client onboarding, status reporting, meeting summaries, sales emails, and support-ticket triage, all activated by asking an agent to use them (see Skills).
External access

Other tools that speak MCP — Claude Desktop, for instance — can call the coordinator's tools through
the app's MCP endpoint. From the coordinator's point of view those requests are handled exactly like
something you typed yourself. Every external call is visible to you in the admin API monitor.

How it behaves

The coordinator is instructed to start with action rather than a long explanation, stay concise, and
if a task took more than a few tool calls, briefly narrate what it did along the way. If a task
looks like something you'll want to repeat, it's expected to point that out — that's a candidate for
becoming a reusable skill.