Most agents fail for the same reason new hires do: nobody told them what the job actually is. Before we write a single line of logic, every Vigil agent gets a one-page job description — reviewed and signed off by the human it reports to.

The five sections

Goal, tools, playbooks, limits and handoff. That’s it. If a section runs longer than half a page, the job is too big — split it into two agents.

role: Inbox Agent
reports_to: ops-lead@company.com
goal: "Reply to 90% of inbound within 5 min"
tools: [gmail, dentrix, stripe]
limits: { refund_max: 250, confidence_min: 0.9 }
handoff: "#ops-escalations with summary + draft"

role: Inbox Agent
reports_to: ops-lead@company.com
goal: "Reply to 90% of inbound within 5 min"
tools: [gmail, dentrix, stripe]
limits: { refund_max: 250, confidence_min: 0.9 }
handoff: "#ops-escalations with summary + draft"

role: Inbox Agent
reports_to: ops-lead@company.com
goal: "Reply to 90% of inbound within 5 min"
tools: [gmail, dentrix, stripe]
limits: { refund_max: 250, confidence_min: 0.9 }
handoff: "#ops-escalations with summary + draft"

Guardrails in plain English

Write limits the way you’d brief a smart temp: “Never promise a delivery date. Never refund more than $250. If a customer mentions a lawyer, stop and hand off.” Plain rules are testable rules.

If you can’t explain the job to a new hire in one page, you can’t explain it to an agent either.

Handoff rules

The best agents know when to stop. Every handoff carries the full thread, the agent’s reasoning and a suggested reply — so the human spends ten seconds, not ten minutes.

Template

• Goal — one sentence, one number.

• Tools — the exact systems it may touch.

• Playbooks — your SOPs, step by step.

• Limits — money, confidence and words it never says.

• Handoff — who takes over, and what they receive.