AI Agent Production Readiness Checklist

Tool scope is minimal — each tool does one thing, with a clear description and typed inputs/outputs

Agent loop has a hard iteration cap (e.g., max 10 steps) to prevent infinite loops

Fallback path defined for when the LLM selects an invalid or unavailable tool

All tool schemas validated with JSON Schema — no free-form unstructured inputs

Context window budget planned — agent doesn't exceed token limits under realistic workloads

Stateful vs. stateless decision made and documented for each agent component

Prompt injection defences in place — user input never directly interpolated into system prompts

Tool permissions use least-privilege — each tool only accesses what it needs

Credentials are never passed through the model context — stored in env vars or secrets manager

Human-in-the-loop approval required for all write/delete/send actions

Output validation — agent outputs are sanitised before being sent to downstream systems

Rate limiting on tool calls to prevent abuse or runaway spending

Audit log of all tool calls with inputs, outputs, timestamps, and user identity

Each tool has a timeout — no tool call can block indefinitely

Retry logic with exponential backoff for transient failures

Idempotency verified for all write tools — safe to retry without duplicating effects

Graceful degradation path — agent returns a useful partial result if a tool fails

Circuit breaker in place for tools calling external APIs

Cancellation signals handled — agent can be stopped mid-run cleanly

Structured logging of every agent run: inputs, tool calls, outputs, latency, token usage

Trace ID propagated across all tool calls in a single agent run

LLM token usage tracked per run and aggregated for cost monitoring

Error rates alerting — spike in tool failures or agent loop exits triggers notification

End-to-end latency benchmarked for p50/p95/p99 under realistic load

Agent output quality sampling — regular human review of random agent outputs

Token budget per run defined and enforced — agent aborts if budget exceeded

Model tier selected based on task complexity — not all tasks need frontier models

Caching implemented for repeated tool calls with identical inputs

Monthly spend limit set with alerts at 50%, 80%, and 100% thresholds

Users are clearly informed when they are interacting with an AI agent, not a human

Agent explains what it is about to do before performing irreversible actions

Undo / rollback available for all agent-initiated changes where possible

Error messages are human-readable — users understand what went wrong and what to do

Feedback mechanism for users to flag incorrect or harmful agent outputs