Reference architecture · 10 minute read
What makes an AI agent production-ready
The architecture, authority and evidence required before an agent performs material work.
- Technology & Data
- Build
- Global
- Build
Decision lens
What must be true before an agent can move from demonstration into accountable work?
A production agent is an operating capability around a model. Workflow, knowledge, tools, identity, authority, evaluation, monitoring and recovery must function together.
Anchor the agent in one measurable workflow
Define the trigger, decisions, handoffs, exception paths and current performance before selecting models or building tool connections.
- Document failure modes
- Define human-agent handoffs
- Set observable acceptance
Treat every tool as delegated authority
System actions need identity, least-authority permissions, input validation, approval rules and auditable outcomes proportionate to their impact and reversibility.
- Allow-list actions
- Separate preparation from execution
- Design stop and recovery
Evaluate the complete system
Test business quality, grounding, authority, reliability, latency, cost and edge cases across the workflow—not only the model response.
- Version test sets
- Record known limitations
- Re-test material changes
Practical checklist
Evidence to bring into the decision.
- 01Workflow specification
- 02Approved knowledge
- 03Tool permission model
- 04Evaluation thresholds
- 05Named service owner
Continue with evidence
Connect the perspective to an operating method.
Production readiness is specific to the workflow, system and accepted risk. No reference architecture proves universal safety or fitness.