State Management in Agent Products

“Agent Basics for Product Managers” part seven
Previous: Agent Autonomy · Series home · Next: Evaluating and Shipping Agent Products

A user says, “Make the second proposal our MVP.” The agent replies, “Done.” Does the system know which object “the second proposal” refers to? Does the user have permission? Did the database write succeed? Will collaborators see the same result?

Without those answers, the interface creates a false sense of completion.

Four kinds of truth

• Business truth: users, requirements, document versions, and permissions.
• Run truth: the current step and what execution is waiting for.
• Conversation truth: what the user and agent said.
• Inference context: what the model saw in one call.

They may reference one another, but they cannot replace one another. A chat log is not durable business state, and memory is not the business database.

A five-layer product architecture

Experience: chat, workspace, progress, diff, approval
Product: users, domain objects, permissions, versions, durable state
Orchestration: agents, workflows, skill routing, pause and resume
Capability: tools, MCP, retrieval, files, external services
Model: inference, generation, routing, limits

Tracing, evaluation, cost, latency, and incident handling cut across every layer.

Separate proposals from writes

Agent analysis → proposal / patch → schema validation
→ user reviews evidence, difference, and impact
→ approve / edit / reject → product layer writes
→ create a version and audit record

A proposal is more than an extra confirmation click. It is an explainable, reversible, conflict-aware object between model advice and business truth.

An effective approval shows what will happen, which objects are affected, the old and new values, the evidence, whether the change is reversible, and what rejection will do. Documents and code need diffs; outgoing messages need recipients and complete content; data changes need impact scope.

Product exercise

Choose one agent action that changes data. Break it into proposal, preview, approval, durable write, verification, and undo. Assign each step to the experience layer, runtime, or product layer.

Conversation supports collaboration. The runtime executes. The product database owns facts. Approval and audit create trust.