Topic guide
A field guide to the architecture around a production agent: explicit state, selective context, bounded tools, deterministic control, evaluation, observability, and a product experience people can trust.
Explore the system
These areas overlap in production. The categories provide a useful path through the existing writing without pretending any one layer solves reliability alone.
Keep the user's current goal explicit and put important transitions in software rather than asking a transcript to carry every decision.
Select bounded evidence for the task instead of treating the context window as a database or every retrieval method as interchangeable.
Make failures observable and reproducible, then turn each important production failure into a regression case or stronger boundary.
Design agent-facing interfaces with clear schemas, boundaries, errors, observability, and contracts that models can use safely.
Treat durable goals, constraints, and application state as explicit data. Conversation history can inform memory, but it should not silently define it.
Build a surrounding product system that handles streaming, persistence, specialized workflows, observability, and failure recovery, with the model as one component.
Architecture in context
Case studies show how these patterns interact with customer needs, product constraints, and measurable outcomes.
AI-assisted workflow
Natural-language intent translated into a validated, testable plugin configuration through an XState workflow.
Read the case studyGrounded research agent
A map-centered agent that coordinates public data, bounded tools, interface state, and traceable answers.
Read the case study