Agent State

Agent state is the agent’s maintained representation of where a task currently stands. It connects separate steps by recording what the agent is trying to achieve, what has already happened, and what still needs attention.

What agent state can contain

State varies by system, but it may include:

• The active goal and subgoals.
• Completed and pending steps.
• Tool calls and their results.
• User approvals or constraints.
• Selected files, records, or resources.
• Retry counts, budgets, and deadlines.
• Errors or unresolved questions.
• A proposed final output.

State may live in application memory, a database, a workflow engine, or a structured object passed between agent steps. Anatomy of an AI Agent shows how state fits with the model, tools, memory, planning, and guardrails.

Observations update state

Suppose a support agent is processing a refund. Its initial state records the order ID and the customer’s request. A tool observation confirms that the order is eligible. The state is then updated to show that eligibility is verified but approval and payment reversal remain pending.

The next action should use this updated state. Without it, the agent may repeat checks, skip required steps, or contradict earlier results.

State is not the same as context

State is the system’s durable task record. Context is the information provided to the model for a specific inference.

The complete state may be too large or sensitive to place in every prompt. A context-building layer can select only the parts needed for the current decision.

State is not the same as long-term memory

State normally describes the active run or workflow. Long-term memory preserves information that may remain useful across separate tasks or conversations.

For example, “awaiting approval for order 123” is task state. “This customer prefers email communication” may be long-term user memory.

Reliable state prevents hidden confusion

Agent systems become difficult to debug when critical progress exists only inside unstructured model messages. Explicit, typed state makes transitions inspectable, supports recovery after interruption, and allows deterministic code to enforce rules before high-impact actions.