Topic reading guide
Production AI agents
An agent needs more than a useful next response. Start with the system boundary, then follow context, actions, recovery and the evidence an operator needs.
Start with the system
AI Agent Architecture
Map the model, runtime and surrounding control systems before choosing a framework or splitting work across agents.
Guide
Context and memory
Designing the Context Layer for Production AI Agents
See how evidence is assembled, filtered by authority and freshness, and kept separate from workflow state.
Essay
Agent Memory vs Execution State
Decide which facts can be reused later and which task state must survive a worker restart.
Guide
Execution and recovery
Designing Long-Running AI Agents That Survive Failures
Read the deeper treatment of state machines, recovery invariants and operational control.
Essay
Control and delegation
Designing Identity and Delegation for Production AI Agents
Keep user, agent and workload identity attached as authority moves through a system.
Essay
An ALLOW Is Not a Permit: Designing Runtime Authorization for AI Agents
Turn an authorization decision into a bounded permit that is checked at the action boundary.
Essay
A Skill Is Not a File. It Is a Deploy: Designing Agent Skills Infrastructure
Treat shared agent skills as versioned deployments with rollout, provenance and revocation.
Essay
Observability
Designing Observability for Production AI Agents
Connect decisions, context, tool effects and outcomes across worker restarts and traces.
Essay
Work through production failures
Explore applied work on execution and recovery. Check the cohort page for its current availability.
Explore the cohort