ICMD 90-Day Agent Launch Kit

Purpose: Ship one agentic workflow to production with measurable value, bounded risk, and predictable cost.

1) Pick the right workflow (Week 1)
- Choose a workflow with a clear “object” (ticket, invoice, alert, PR) and a measurable outcome.
- Define acceptance tests in plain language (e.g., “Resolve 75% of L1 tickets without escalation”).
- Estimate value: minutes saved per task × tasks/week × blended hourly cost.
- Identify irreversible actions (send, delete, refund, deploy) and mark them “confirmation required.”

2) Define boundaries and policies (Weeks 1–2)
- Decide agent level: Suggest / Act with confirmation / Autonomous within policies.
- Write an allowlist of tools (3–6 max) and a blocklist of dangerous tools.
- Define permission model: OAuth scopes, service account, and any break-glass escalation path.
- Specify PII rules: fields to redact, storage retention, and logging access controls.

3) Build tool wrappers and observability (Weeks 2–6)
- Implement typed tool interfaces (strict inputs/outputs) and validation.
- Add rate limits, timeouts, retries, and idempotency keys to tool calls.
- Log: user intent, plan, tool traces, retrieved evidence, final action, and timestamps.
- Add replay: ability to re-run a failed task deterministically with the same inputs.

4) Evaluation and rollout plan (Weeks 4–10)
- Create a golden set of 50–200 real tasks with expected outcomes.
- Define metrics: Task Success Rate (TSR), cost per successful task, intervention rate, rollback rate, evidence coverage.
- Establish launch gates (example): TSR ≥ 75%, rollback ≤ 3%, cost/success within tier ceiling for 14 days.
- Canary release: 1–5% traffic; compare against control cohort.

5) Packaging and pricing (Weeks 8–12)
- Pick pricing axis tied to objects (per ticket/invoice/alert) or autonomy tier (Suggest vs Act).
- Publish cost ceilings per workspace/tenant; add admin controls to cap spend.
- Create an “agent report” dashboard: time saved, completion rate, errors/rollbacks by team.

6) Production readiness (Weeks 10–13)
- Incident runbook: how to disable tool access, roll back actions, and notify affected users.
- Security review: audit log integrity, least privilege, data retention, and vendor risk.
- Customer enablement: admin docs, policy templates, and a “what the agent can/can’t do” page.

If you complete these steps, you’ll have a shippable agent: narrow, measurable, governed, and economically predictable—ready to expand to the next workflow.