A long-running agent loses track of its goal after many turns. What is the best context strategy?
Why D is the answer
Keep a short state record of goals and decisions, resent each turn
Curated state survives where a full transcript does not. Every token in context competes for attention and budget.
In detail
- Curated state survives where a full transcript does not.
- Keep goals and decisions in a short record, and resend it on every turn.
- Update the record as decisions change.
Use case
An illustrative scenario, not a real incident.
A migration agent forgets its goal after hours of work. A short state record of the goal and decisions, sent each turn, keeps it on course.
See it in different ways
stateDiagram-v2 [*] --> SingleCall SingleCall --> Workflow: steps are known Workflow --> Agent: path depends on findings Agent --> MultiAgent: evals show overflow or specialist failures MultiAgent --> [*]
mindmap
root(("State record"))
n1["Goal"]
n2["What done means"]
n3["Decisions"]
n4["Choices so far"]
n5["Next"]
n6["Remaining steps"]Key point: Curate state, rather than keeping everything.
Why the other options are wrong
- A. Keep every tool result in full, so that nothing can ever be forgotten
- Keeping everything crowds out the goal, and it grows until it hits the limit.
- B. Double max_tokens at each turn as the conversation gradually grows
- max_tokens limits the reply, not what the model holds in context.
- C. Remind the model in capital letters to remember what it is working on
- Capital letters cannot restore lost information.