FirstMate: agent crews, durable evidence and delivery authority
When to choose FirstMate instead of one coding agent or a workflow engine
Select by task independence, human decision points and durable-state requirements rather than treating multi-agent execution as a universal upgrade.
What you will learn
- Choose based on the coordination problem
- Compare the work boundary, not a feature count
- Adopt with an exit criterion
Before you start
- Git worktrees and pull-request basics
- Understanding of terminal agents and credential scope
Define a delivery contract and inspect state and authority evidence without claiming untested runtime guarantees.
Key takeaways
- Parallel coordination is useful only when the workload needs it.
- Deterministic validation and agent investigation complement each other.
- Evaluate one concrete harness/backend combination.
Choose based on the coordination problem
A single bounded change often needs one agent and ordinary version control. FirstMate becomes more relevant when several independent investigations or changes need persistent supervision and a single human liaison. Its extra state and backend setup are justified only if that coordination burden is real in your workflow.
A deterministic CI workflow addresses a different problem: executing known checks and transitions from explicit inputs. An agent crew interprets requests and investigates uncertain work. Keep repeatable validation in CI where appropriate, even when agents prepare the change; natural-language coordination is not a replacement for reproducible checks.
Compare the work boundary, not a feature count
List what each task may read, write and deliver. If tasks share an unstable interface, splitting them can increase reconciliation work despite isolated checkouts. A scout may be the better first step when the implementation boundary is unknown. If you cannot define acceptance evidence, adding a supervisor will not make completion objectively measurable.
Evaluate the selected backend and harness combination, not a generic claim that every integration behaves identically. At the pinned revision backend coverage and semantic state support differ. A visible session is useful for inspection, but it does not automatically give every adapter the same recovery guarantees.
Adopt with an exit criterion
Trial two independent tasks over disposable repositories and compare accepted outcomes, human attention and recovery complexity with your current process. Reject the configuration if it silently widens authority, cannot explain task state or leaves work without an accessible deliverable. This proposed trial is not a comparative benchmark we executed.
Keep a clear way to stop spawning new work while preserving current task state and unmerged changes. If the fleet costs more attention than it saves, a smaller setup may be preferable. The recommendation is conditional on your observed workload, not on the project’s position in a weekly popularity chart.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Compare against your existing single-agent process.
- 2
Identify independent tasks and shared integration points.
- 3
Define acceptance and authority failure criteria.
- 4
Adopt only if the trial reduces meaningful coordination burden.
Copy-ready example
{
"selectionWorksheet": true,
"independentTasks": 2,
"sharedApiChanges": false,
"acceptanceDefined": true,
"backendRecoveryTested": false,
"adoptionDecision": "pending trial"
}Frequently asked questions
Should every coding task use a crew?
No. A single bounded change may not justify fleet state and supervision overhead.
Can FirstMate replace CI?
Not as a general rule; reproducible checks should remain explicit even when agents coordinate the work.
Sources
- FirstMate / README.mdSource checked 2026-09-14
- FirstMate / docs/architecture.mdSource checked 2026-09-14