Agent Workflow Readiness Checklist (2026)

Use this checklist to go from “agent demo” to a production workflow you can sell, support, and govern.

1) Workflow Selection (Problem Fit)
- Name the workflow in one verb phrase (e.g., “triage inbound support tickets”).
- Definition of Done is binary or strongly verifiable (artifact created, field updated, task closed).
- Baseline metrics captured today: cycle time, volume/week, error rate, cost per unit.
- Economic target set: e.g., reduce cycle time by 25% in 60 days, or cut escalations by 30%.

2) Scope and Constraints (Safety Fit)
- Action space is small (start with 3–7 tools/actions).
- Inputs/outputs have schemas (JSON or typed objects) and validation rules.
- Clear “stop conditions” exist (max steps, max retries, max spend, timeouts).
- Escalation path defined (who gets notified, where the queue lives, what context is attached).

3) Risk Tiering (Autonomy Fit)
- Assign a tier:
  Tier 0 Read-only, Tier 1 Drafts, Tier 2 Internal writes, Tier 3 External actions, Tier 4 Money/privilege.
- For Tier 2+, confirm rollback/undo strategy (single record + batch revert).
- For Tier 3+, implement approvals, allowlists, and immutable audit logs.
- For Tier 4, enforce two-person rule and staged rollout by policy.

4) Observability (Operability Fit)
- Tracing is enabled end-to-end (request → plan → tool calls → validations → outcome).
- You log structured events (tool name, params hash, success/failure, retries) with redaction.
- Dashboards exist for: completion rate, escalation rate, tool-call success, p95 latency, cost per task.
- You can replay a task from a trace to reproduce failures.

5) Evaluation (Quality Fit)
- Define what “failure” means for this workflow (wrong customer, wrong amount, wrong permission).
- Build a golden set: at least 50–200 historical cases with expected outcomes.
- Add automated checks: schema validation, policy checks, dry-run execution where possible.
- Set launch gates: e.g., >90% completion on golden set and <5% critical errors.

6) Packaging & Pricing (Business Fit)
- Name the unit of value (resolved ticket, processed invoice, completed review).
- Choose a predictable plan: base tier includes X units/month; overages priced per unit.
- Include spend controls: caps, alerts at 50/80/100%, per-workflow quotas.
- Document ROI narrative with numbers from pilots (before/after cycle time, throughput, quality).

7) Launch Plan (Adoption Fit)
- Start with “suggest/approve” mode before full autonomy.
- Ship a review inbox UI for approvals and escalations.
- Create playbooks for top 10 edge cases; update weekly from operator feedback.
- Define the 30-day target: expand to a second workflow only after meeting reliability KPIs.

If you can’t check at least 80% of the items above, don’t scale autonomy. Tighten the workflow, reduce actions, and instrument first.