Agent playbooks

These playbooks assume you have run sx project connect and the agent can see the sx MCP tools and project skill. Replace every <PLACEHOLDER> before sending.

Orient to a project

Inspect the connected sx project. Do not mutate anything. Inventory the exact
tasks, datasets, applications, active operations, deployments, and unresolved refusals
relevant to <OUTCOME>. Use platform tools only; do not ask for repository access or raw
service URLs. Return facts first, then the smallest evidence-producing next action.

Good output names concrete resource and operation references. It does not produce a generic platform summary.

Curate training data

Build an immutable training selection for task <TASK_REF> from datasets <DATASETS>.
First inspect task coverage and search whole segments for <BEHAVIOR>. You may read and
freeze a selection; do not capture, annotate, export, delete, or start compute. Explain
the inclusion rule, list exact selected members, call experience.selection.snapshot once,
and return the snapshot identity plus missing coverage.

Expected path: experience.task_map → experience.search_segments → review → experience.selection.snapshot.

Evaluate a candidate

Evaluate application <APPLICATION_REF> against task <TASK_REF> with seed <SEED>.
Inspect both exact resources and compatibility before mutation. You may submit exactly
one Worlds evaluation. Save the operation reference immediately, follow it with bounded
operation updates, and report success only from operation.result. Return the terminal
evaluation, failure strata, and evidence references. Do not train, promote, or touch hardware.

Expected path: autonomy.application + experience.task → worlds.evaluate → operation.updates → operation.result.

Run a bounded improvement loop

Improve task <TASK_REF> for at most <ROUNDS> rounds and <SECONDS> seconds. First summarize
the existing data, candidate, and evaluation evidence. Ask before starting if the stated
bounds or grants are insufficient. Once authorized, call autonomy.improve once, persist
the operation reference, and monitor it. Do not widen scope, increase spend, start physical
research, or deploy. Return each round's hypothesis, evidence, decision, and terminal result.

The round and time bounds are hard authority. Losing the agent connection is not cancellation.

Prepare a deployment decision

Prepare a promotion brief for application <APPLICATION_REF> to group <GROUP>.
Do not deploy. Verify the exact application, task evaluations, target qualification,
capability approvals, supervision requirements, current healthy deployment, and rollback
target. Put contradictions and missing proof first. End with one of: READY FOR HUMAN
DECISION or NOT READY, followed by the exact reason.

Open the Fleet Console for the accountable promotion decision and live supervision.

Triage a stuck operation

Diagnose operation <OPERATION_REF>. Do not submit replacement work. Read its current status,
then page ordered updates from the last known cursor <CURSOR>. If terminal, fetch the result.
If failed or refused, name the owner, typed code, effect, and documented recovery. If history
has a gap, use the owner's latest authoritative receipt. Do not cancel without my approval.

An agent may complete routine work inside a clear project, task, budget, and time boundary. New spend, held-out-data use, physical motion, cancellation, promotion, and safety-policy changes need explicit authority.

sx docs