Workspaces and Specialists: Why We Spawn a Team Instead of One Agent
A SE Studio workspace is a way of working. Inside it you spawn specialist agents — each scoped to one job — because focused specialists outperform a single generalist.
Point one capable agent at a whole feature and it will optimize for the next answer. It writes the code, half-writes the tests, skips the security implication it was not asked about, and declares victory. The work is not wrong, exactly. It is unscoped. We kept seeing the same fix: give the work to a team, not a generalist.
A workspace is a way of working
The first choice in SE Studio is not which agent — it is which workspace. A workspace carries a mode: a standard local workspace for hands-on coding, a Sprint Engine workspace for planned multi-role sprints, a Switchboard workspace for issue intake. Each mode brings its own surface and its own default team, so you are picking a way of working before you pick who does the work.
Specialists, scoped to one job
Specialists in a workspace
Each specialist gets a narrow brief, its own session, and its own context. The architect plans and does not implement. The security reviewer audits and does not ship features. That scoping is what keeps an agent from wandering: a tester asked only to prove the change works will write the failing case the developer skipped.
- architect — turns the brief into a task graph with dependencies.
- developer — implements a specific task in an isolated worktree.
- frontend — builds UI against the existing component conventions.
- security — audits auth, tenancy, and sensitive mutations.
- tester — runs and records validation before review.
- reviewer — leaves a verdict and a residual-risk note.
Why specialists beat a generalist
This is not just intuition. Across the local swarm runs we reviewed — 26 boards and 482 tracked tasks — the agent signal was strongest exactly where scoping was clear. Average task clarity sat around 94 percent, confidence around 92 percent, and role fit around 98 percent. When an agent is handed one job with a clear boundary and acceptance criteria, it executes well.
A generalist loses that. It has to hold the entire feature in context, juggle competing concerns, and decide on its own which corners are safe to cut. A specialist with a small remit keeps its context tight and its output sharp — and the workspace gives those specialists a shared place to hand off.
Spin up a workspace per way of working, then spawn specialists per job. The team is the unit, not the agent.