hard20 pts · first answer countscontext-windowagents

Long agent run fills context

A coding agent works for hours; tool outputs pile up until the context is full, and quality drops well before that. What is a good strategy?

A.Restart from scratch whenever the context is fullB.Pick the model with the biggest window and keep everythingC.Delete the system prompt to make roomD.Compact the history: summarise or drop old tool outputs, keep key decisions and state, and store notes in external files the agent can read back

Discussion 0

Join free to post in the discussion.

Long agent run fills context

Discussion 0

Related questions