Agent State
Everything the agent is currently holding in mind about the task — the goal, progress so far, and what it knows right now.
Think of it like
The sticky notes on your desk mid-project: what you’re doing, what’s done, what you found out this morning.
Example
The agent’s state includes the original request, the three sources already gathered, and a note that the fourth link was broken.
How it actually works
State is the evolving bundle carried across loop turns — goal, plan, intermediate results, memory — that makes an agent more than a stateless single response. Managing it is a central challenge: too much and it overflows the context window; too little and the agent forgets what it was doing. Much of agent engineering is deciding what to keep in state, what to summarize, and what to offload to memory.
For product teams
The thing that lets an agent hold a thread across many steps — and the thing that quietly breaks when it can’t.
For engineers
The maintained set of goal, plan, and intermediate results carried across steps, bounded by context limits.
Related
- Handoff — What gets passed when control moves between agents.
- Agent Memory — The longer-term store state spills into.
- Trajectory — The path this state traces over a run.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome