ICMD Agent Workflow PRD + Launch Readiness Checklist (2026)

Use this template to ship one production-grade agent workflow in 2–6 weeks.

1) Workflow Definition (Versioned)
- Workflow name + version (e.g., invoice_reconciliation_v2)
- Target user persona (job title, team size, tools they use)
- Job-to-be-done statement (one sentence)
- In scope: exact task boundaries (what the agent will do)
- Out of scope: explicit exclusions (what it will never do)
- Inputs: required fields + optional fields + allowed formats
- Outputs: required artifacts (records created, fields updated, files produced)
- Reversibility: how to undo actions (rollback plan)

2) Success Metrics (Pick 3–5)
- Verified Completion Rate (VCR): % runs that pass checks
- Time-to-outcome: median + p95 minutes per run
- Human Minutes Saved (HMS): baseline minus post-agent time
- Escalation rate: % runs requiring human review
- Cost per verified outcome: total run cost / verified completions

3) Permissions & Guardrails
- Least-authority role definition (what systems, read vs write)
- Approval points (what requires human confirm)
- Policy checks (rules engine, thresholds, required citations)
- Data handling: PII redaction, retention window, export controls
- Audit log fields: run_id, actor, tools called, data sources, changes made

4) Evaluation Plan (Before Launch)
- Golden set: 50–300 representative cases (label source, freshness date)
- Scoring dimensions: correctness, groundedness/citations, format compliance
- Regression gates: define “no-ship” thresholds (e.g., VCR drop >2%)
- Drift monitoring cadence: weekly review + monthly refresh of golden set

5) Observability & Ops
- Run dashboard: VCR, error rate, latency p95, cost/run, retries
- Incident playbook: how to pause workflow, roll back, notify customers
- Support workflow: how CS accesses run history and exports evidence

6) Pricing & Packaging
- Value hypothesis: hours saved or revenue protected per month
- Pricing model: platform fee + per verified outcome OR tiered usage buckets
- Limiters: max runs/day, max write actions, premium for high-stakes modes
- ROI report: monthly customer summary metrics to defend renewal

7) Launch Plan (2-stage)
- Stage 1 (Beta): read-only or reversible actions; heavy logging; opt-in users
- Stage 2 (GA): expanded permissions; published SLA; documented controls
- Post-launch: weekly eval review; ship workflow versioning + changelog

If you can’t answer one item above with a specific number, policy, or owner, you’re not ready to scale the workflow—yet.