HumanBee

Guide

AI agent production checklist

A practical checklist for shipping production AI agents: durability, tools, authorization, HITL, integrations, and ops.

Use this checklist when moving from a demo agent to something your backend team can operate. HumanBee uses the same themes in architecture and delivery work.

1. Outcome and boundaries

Define the business outcome, the first path only, and what the agent must never do unsupervised.

  • Single measurable acceptance criteria for the pilot
  • Explicit out-of-scope actions
  • Owner for the agent in your org

2. Tools and integrations

Inventory tools by side effect. Prefer typed adapters against your APIs over generic connectors.

  • Read vs write vs irreversible classification
  • Auth and tenancy matching your estate
  • Error and compensation paths for writes

3. Authorization and HITL

Default-deny protected tools. Human approval for irreversible actions until trust is earned.

  • Deterministic policy — not LLM judgment alone
  • Suspend/resume for approvals
  • Audit trail of who approved what

4. Durability and recovery

Runs must survive restarts and deploys. Partial failure should be visible and recoverable.

5. Observability and handover

Operators need timelines, failures, and runbooks. Document how your team extends the system.

6. Privacy topology

Document where prompts, logs, embeddings, and model traffic go — especially for private or regulated deployments.

FAQ

Is this checklist HazelJS-specific?

No. The themes apply across stacks. HazelJS is one Agent OS option when TypeScript fits.

Can HumanBee run this as a workshop?

Yes — architecture workshops and enablement sessions are available. Start via consultation.