Orchestrating a fleet

A single agent is a run; several of them make a fleet you watch from one dashboard. Precursor lets agents park until triggered, budget and retry themselves, and fire on schedules or external events — with a fleet-wide concurrency cap over the lot.

Chaining agents into a pipeline

Sequencing one agent after another — research → draft → review — is owned by Workflows, a reusable coordinator that runs a series of agents in order and passes each step's output to the next. Individual agents stay independent; a workflow supplies the chaining.

An agent is a definition; each start is a run

An agent is a reusable definition — a title, an objective, a model, a role and its capability defaults. Every time something starts it, Precursor opens a separate run: its own status, prompt, transcript, artifacts, token meter and Copilot SDK session.

That split is what makes an agent safe to share. Two workflows can point a step at the same agent and both be in flight at once — each drives its own run, so neither sees the other's status, wipes the other's artifacts, or is charged for the other's tokens.

A run records what triggered it (manual, workflow, schedule, webhook, fleet, or retry / replay), so the history reads as an audit trail. Runs are kept, never garbage-collected, and the insights sidebar carries a Runs rail listing the recent ones with the current one highlighted.

The transcript shows the latest run by default, and that rail doubles as a filter: click any run to read that execution alone, or All runs to stitch the history back together. This matters most for a shared agent, where two workflows driving it at once would otherwise produce one interleaved archive.

Park an agent until a trigger

Not every agent should run the moment you create it. The composer's Create parked (don't run yet) toggle arms an agent in the waiting state instead of starting it — fully configured but idle until its webhook fires, you hit Start now, or a workflow step reaches it. A parked agent is never picked up by the orphan sweep, so you can pre-stage a run and fire it when ready.

Budgets & the concurrency governor

• Concurrency governor. A fleet-wide cap (agents_max_concurrent, default 3) limits how many agents run at once; the rest queue and are released as slots free up, so a burst of new agents can't stampede the runtime.
• Token budget (per agent). Once an agent's cumulative usage crosses its budget, it parks itself in the inbox for approval instead of spending more. Leave it blank for unlimited. The budget is cumulative on the definition: spend from everyrun counts against it, so restarting doesn't hand it a fresh allowance.

Retry & auto-recovery

Set max retries on an agent and a failed run is automatically retried with exponential backoff (agents_retry_backoff_seconds × 2ⁿ) before it's finally marked failed, so a transient error recovers itself without you noticing.

Blueprints (reusable templates)

A blueprint saves a task prompt plus a governance profile (model, approval policy, token budget, retries) as a reusable template. Manage them under Settings → Agents → Blueprints; hit Run on one to stamp out a fresh agent with the same profile — a repeatable job kept one click away instead of retyped.

Event triggers (webhooks)

An agent can be re-run by an external event. Add a webhook from the insights sidebar to mint a unique URL (POST /api/agents/hooks/{token}); calling it re-runs that agent. Bad or disabled tokens 404, so there's no enumeration.

Scheduling agents

An agent can carry its own recurrence so it re-runs its stored task on a cadence. Each due tick either replays the task on a fresh transcript (clear_context) or sends a follow-up in the existing conversation. A run is skipped (not failed) while the agent is mid-turn, waiting on you, archived, or task-less.

You can also drive an agent from a scheduled topic with slash directives:

• /agent <uuid> /clear <follow-up> — reset the transcript and freshen the blackboard, then send a follow-up.
• /agent <uuid> /run [extra] — reset, then replay the agent's own task prompt plus an optional one-off extra.

See the scheduler for how these directives — and /guard gating — fit together.

The unified inbox

Everything waiting on you — parked approvals, raised Needs input questions, and budget parks — is collected into one inbox strip across the top of the dashboard, each chip deep-linking to its agent. The dashboard header also shows a fleet rollup: how many agents are running vs queued, and total tokens consumed against the sum of budgets.