Agent Launcher Orchestrator

Agent Launcher Orchestrator

When to use

Use to plan or launch a managed agent that acts repeatedly through tools under a named owner's account.

Boundaries

Use a prompt-design skill for one-off text behavior and Agent Loops for an internal loop without a managed launch. This skill does not grant deployment permission.

Inputs

Collect the job to automate, allowed tools and accounts, data classification, owner, spending ceiling, frequency, required approvals, success measure, and incident contact.

Method

1. Convert the goal into bounded tasks with explicit inputs, permitted actions, irreversible effects, and stop conditions. Map every tool call to the minimum permission it needs.
2. Define an evaluation set with normal, ambiguous, adversarial, and tool-failure cases. Include a rubric for outcome quality and policy compliance.
3. Run a supervised pilot. Record traces, costs, failures, human interventions, and false success reports. Launch recurrence only after the owner reviews the evidence and monitoring plan.

Decision rules

• If a tool can spend money, publish, message others, or modify production data, require an explicit action boundary and approval gate.
• If the agent cannot detect a failed or partial action, stop the rollout until observability and recovery are designed.

Verification

• The agent finishes a representative task within budget and permission limits.
• It halts or escalates on missing context, denied permissions, and repeated tool failure.
• The owner can reconstruct what it did and disable recurring runs promptly.

Stop and handoff

Do not launch recurring actions without owner approval, bounded permissions, a spend limit, observable failure signals, and a disable path. Record the blocker and who can resolve it.

Output

Return a launch packet with task contract, tool-permission matrix, evaluation cases, pilot evidence, budget, monitoring, rollback, and named owner.