Hand a job to one agent and watch it grow a team
Start a single overseer, describe an outcome instead of picking a roster, answer it when it asks, and see why nothing sits between you and the workers it starts.Outcome and prerequisites
You will start one agent, give it a goal rather than a task list, and let it decide whether the work needs a team at all. Along the way you will answer a question it cannot answer itself, and message a worker it started without going through it — which is the difference between this and every other way of running agents.
- The
synapseCLI installed and onPATH, with the mesh on:synapse settings mesh on. - Claude Code installed and signed in. The agents run under whatever account you already have — Synapse holds no model credential and makes no model call.
- A checkout you can afford to have edited. Agents here touch real files.
- About thirty minutes.
If the mesh is new to you, read Run a team of agents and drive it yourself first. This tutorial is the same machinery with one agent instead of four.
-
Start one agent, not a roster
The other teams ask you to choose a shape before you understand the job.
overseeris one member:cd ~/code/your-project synapse relay team show overseer synapse mux --team overseerTwo things happened. An agent started in the background under the
overseerrole, and you joined the mesh under your login name. Check both:/agentsYou should see two rows: yourself, marked as a person, and
overseer. The mark matters — every agent on the mesh is told to ask a human row questions and never delegate work to one. -
Describe an outcome
Say what you want to be true, not which steps to run:
@overseer the release notes for the next version are missing. Work out what changed since the last tag and write them.Watch what it does before it does anything:
/agentsIts note is the interesting part. An overseer's brief tells it to weigh doing the work itself against splitting it, because every worker it starts is a separate session on the same account you are paying for. A small job stays one agent. A job with two genuinely parallel halves gets a second.
-
Answer it when it asks
Sooner or later it will reach something only you can settle — which version number, whether a change is worth mentioning, whether to touch a file it is unsure about. It will send you the question and report itself blocked:
/agentsA row reading
blockedwith a note is an agent waiting on you. Reply to it like anything else:@overseer call it 2.4.0, and leave the vendored files out entirelyThis is the whole reason a person belongs on the roster. A headless worker runs with its permission prompts bypassed, so without somebody to address, its only option at a fork is to guess.
-
Reach past it
If the overseer started workers, they are on the roster too — and you can address them directly:
/agents /log notes@notes skip the dependency bumps, they are noiseNothing was relayed. You did not interrupt the overseer's context to correct one of its workers, and it did not have to spend a turn passing the message along.
-
Do the same from the app
Open Synapse and click Console. It is the same seat: a transcript of everything said, what the mesh is doing, the roster, and a box to type in. Opening it registers you under the same login name, so an agent looking for somebody to ask finds one row rather than two.
Click an agent in the roster to aim a bare line at it, or use
@nameexactly as in the terminal. The grammar is one piece of code shared by both surfaces. -
Bound the fan-out
How many agents a machine can usefully carry is a fact about the machine:
synapse settings workers 4 synapse settings show | grep workersA supervisor already running picks this up on its next spawn. The number is clamped to a ceiling in code as well as refused on the way in, so a mistyped digit cannot buy an unbounded fleet — the setting is a preference, and the ceiling is the promise.
-
Clean up
/workers /kill notes /quitLeaving takes the workers it started with it, and drops your roster row so no agent goes on addressing questions to somebody who is not there.
What to read next
Run a team of agents and drive it yourself is the same mesh with a roster you choose. Let your agents write skills turns a procedure one of these sessions worked out into something the next one starts with.