orchestrate is the one agent that orchestrates, a workflow is a plan it is started on, and the live bench compares it with build on the same tasks¶
Decision 888 · accepted · 2026-10-12 · issue #554 · supersedes 850
The decision¶
- One agent orchestrates:
orchestrate. Theworkflowbuilt-in is gone.orchestratetakes its model,expensive, which is the default model unless a person names a dearer one (models.expensive), so nothing changes for anyone who has not, and its paragraph: a prompt that is a numbered step list is a workflow, the steps are the plan,[you]is its own step,parallelsteps may go out together, and no step, least of all a test or verify step, is dropped without saying why. The rendered plan says the rest, as it did. - A workflow is a plan
orchestrateis started on.Troupe.Workflow.agent/0names the runner, andsession.createwithworkflowstarts it onTroupe.Workflow.plan/2's prompt. A client that namesprofile: "workflow", the old runner's name, getsorchestrate(Troupe.Workflow.agent/1); one that names another agent runs the plan there, as before. Worktrees, approvals, budgets and resume apply unchanged, since running a workflow is still starting a session (Decision 648). /workflowis a built-in command in the harness's table and the terminal client's, since no agent row carries it any more; the terminal starts the branch onorchestrate, in a worktree of its own, its window stillworkflow-<n>. The desktop app starts no branch and lists it as one it cannot run.- What this replaces. Decision 850's choice of two profiles (the rest of 850 stands),
TUI Decision 87's "the
workflowagent" (its reasoning holds fororchestrate: an expensive orchestrator that cannot write and owns only the decisions), and TUI Decision 104's "theworkflowagent's first input" (noworchestrate's). - Not chosen. Two profiles with a stated boundary: the step list was the only thing the
workflowagent had thatorchestratedid not, and a list is data the plan renders, not a second agent to keep in step.workflowas the name of the one agent:orchestrateis the name #554 and Decision 850 gave the profile people choose, andworkflownames the plan.orchestratehanding a multi-step task to a workflow through a tool: a second session for what a task list already does. - A repository's own
.troupe/agents/workflow.mdis an ordinary primary now, which/workflowdoes not run; one that changed how workflows run copiesorchestrate.mdinstead.
The bench¶
- The same outcome at two and a half to three times the cost. Every run of both agents
got the work done and held every check.
orchestratecost 2.5, 2.9 and 3.1 times whatbuilddid, made twice the model calls (2.5 times onfollow_steps), and took 2.4 to 2.8 times as long. - So
buildstays the default andorchestrateis suggested nowhere. On these tasks it buys nothing a person can see. What the bench does not have is a task large enough thatbuild's own context grows past what a few briefs cost, or one where a reviewer catches what the writer missed; every task here takesbuildthree or four calls.
Which to use¶
build for a change you would make yourself, which on the bench is every one of them;
/workflow when a repository has written down the steps a change goes through (a release,
a migration) and wants each owned by a subagent, with a reviewer at the end, in a worktree
of its own; orchestrate on your own words when you want that shape without a declared
list, knowing that on small tasks it costs about three times as much. The terminal
client's README says the same, with these numbers.
Governs: apps/troupe_core/lib/troupe/workflow.ex, apps/troupe_core/lib/troupe/commands.ex, apps/troupe_gateway/lib/troupe/gateway/dispatch.ex, apps/troupe_core/lib/troupe/bench/live_scenarios.ex, clients/tui/lib/troupe/client/daemon.ex, clients/tui/lib/troupe/ui/tui/server.ex