How Agent Runs Work

“Agent Basics for Product Managers” part three
Previous: When to Use an Agent · Series home · Next: Context, Sessions, and Memory

When the interface shows only a chat box, it can look as if the model does everything. In reality, the model repeatedly proposes a next step. The agent runtime makes the task operate.

Think of the runtime as the agent's operating system. It decides what the model can see and call, where actions execute, how the process is recorded, and when it must stop.

A complete run

Chat / API / scheduled trigger
  ↓
Identify the user, goal, and risk
  ↓
Build context from rules, sources, memory, and tool definitions
  ↓
Model chooses: answer / tool / question / approval request
  ↓
Runtime validates permission and parameters, then executes
  ↓
Return the result or error to the model
  ↺
Complete, fail, reach a limit, or wait for a human
  ↓
Persist state, outputs, and the audit trail

The repeated model–tool–result cycle is the agent loop.

Runtime responsibilities

A production runtime must handle entry identity, context selection, loop limits, tool execution, resumable state, memory, permissions, observability, and evaluation. Product managers do not need to implement each module, but the product requirements and acceptance criteria must acknowledge them.

Context building often matters more than changing models

The model does not automatically know company data, project rules, or a document that changed moments ago. The runtime selects relevant information for every call. Loading everything creates noise, delay, stale conflicts, privacy risk, and results that are difficult to explain.

The same model can behave very differently across products because the context, tools, and feedback loop differ.

Stopping is more important than looping

A controlled agent knows when the goal is satisfied, when it needs clarification, how to respond to a failed tool, when time or cost limits have been reached, when approval is required, and when safe progress is impossible.

“Keep going until it is good” defines neither success nor failure.

Product exercise

Draw one complete run for an agent feature. Mark every read, model judgment, tool action, state write, approval, and stopping condition. Any node labeled only “AI handles it” needs another level of definition.

The model proposes what might happen next. The runtime decides whether, where, and how it happens—and when it must stop.