Six Designs for Production AI Agents: Authority, State, Execution, Observation, Evaluation and Recovery
Control how failures propagate into external actions, beyond model response quality.
- AI
- Identity
- Reliability
Architecture Notes
Explore system boundaries, dependencies, technology choices and structures that accommodate change. Record why a design is chosen, which constraints shape the decision and which trade-offs remain.
10 Articles
Control how failures propagate into external actions, beyond model response quality.
Keeping data local leaves capacity, updates, monitoring and recovery as operational responsibilities.
Contain an irreplaceable external API in an adapter and separate the application contract.
Authentication at entry does not by itself control lateral movement or compromise of recovery systems.
Assume compromise and define responsibility for detection, containment and trusted recovery.
Connect collection upgrades to signals that explain failures and declarative change management.
Define safe callable contracts and workflow boundaries that can return decisions to people.
Identify workload and observable improvement conditions before selecting technology.
Make criteria, exceptions and responsibility inspectable before translating human judgment into prompts.
Assess production ownership and sustained evaluation separately from prototype accuracy.
No articles match this topic. Please choose another topic.