Agent session

Agentic AIArchitecture and orchestrationPublished By Simon Budziak

An agent session is one bounded period of interaction and execution with its own messages, tool results, state, and identifier. Sessions make work resumable and auditable by separating one run from persistent memory, while still allowing a new run to load approved context from earlier activity.

OpenAI Agents SDK provides the primary reference used for this definition and its production boundaries.

How does agent session work in production?

The session stores conversation and tool events for one run. Agent memory holds information meant to outlive it, while an agent checkpoint records resumable execution state. Session history is evidence, not permanent memory.

When does agent session matter?

Create a new session when ownership, permissions, or the task boundary changes. A long-running agent may span several sessions, and an agent audit trail links them. Stable identifiers matter more than keeping one context window alive.

Frequently asked questions

What is agent session used for?

Create a new session when ownership, permissions, or the task boundary changes. A long-running agent may span several sessions, and an agent audit trail links them.

Should every agent conversation use the same session?

No. Separate sessions keep permissions, tasks, and evidence from bleeding into unrelated work.

Summarize this page with

See this working in a system we built