Agent Feature Production Readiness Checklist (2026 Edition) Use this before shipping an agent that can take actions (write data, trigger workflows, move money, or deploy code). Target: it can run unattended inside clearly defined boundaries. 1) Scope & Success Criteria - Describe the workflow in one sentence (example: “triage inbound support tickets and draft replies”). - Define a measurable target tied to the workflow’s outcome. - Write explicit non-goals (what the agent must never do). 2) Capabilities & Permissions - Implement capability-based access (examples: issue_refund, update_crm_field, open_pr). - Enforce least-privilege OAuth scopes and short-lived tokens. - Separate sandbox vs production credentials and environments. - Require idempotency keys for any write action. 3) Cost, Time, and Loop Controls - Set per-run ceilings (max tokens, max tool calls, max wall-clock time). - Set tenant/workspace budgets and rate limits. - Add stop conditions and maximum iteration counts. - Attribute cost per run: model spend, tool/API spend, retries. 4) Tool Contracts & Reliability - Define tool schemas strictly; validate inputs and outputs. - Version tools and prompts; ship changes behind flags. - Add circuit breakers and retries with exponential backoff. - Track tool success rate, tail latency, and rate-limit errors. 5) Observability & Audit Trail (“Flight Recorder”) - Store structured traces: inputs, retrieved sources (with IDs), decisions, tool parameters, outputs. - Redact or minimize sensitive payloads; document retention rules. - Provide a user-visible execution timeline and a clear outcome artifact. 6) Human Oversight & Reversibility - Use diff-first review for state changes (field diffs, PR diffs, invoice diffs). - Configure tiered approvals (by risk level, dollar amount, environment). - Provide undo/rollback where possible (revert PR, compensating transaction). - Track override and rollback rates. 7) Evaluation & Release Safety - Build a replay harness to rerun past jobs against new models/policies. - Maintain a small golden set plus ongoing production sampling. - Alert on regressions: completion rate, escalation rate, policy denials, error categories. - Run controlled tests for thresholds (confidence gating vs review load). 8) Rollout Plan - Phase 1: Draft-only suggestions. - Phase 2: Assisted execution (reversible writes) plus approvals. - Phase 3: Delegated execution inside strict scopes (by domain, environment, policy limits). - Phase 4: Managed autonomy with customer-configurable policies and service commitments. If you can’t answer “Who approved this?”, “What changed?”, “How do we undo it?”, and “What did it cost?” you’re not ready to ship.