More context is not better. Past a threshold, accuracy falls off a cliff (context rot).
Don't dump whole files - read the relevant function/section, not the 2000-line module.
Summarize, don't accumulate - replace a 200K-token transcript with a 4K recap of the load-bearing facts.
Drop dead context - once a sub-task is done, its detail leaves the window. Keep the decision, drop the trace.
State on disk, not in context - progress goes to a file the next turn re-reads, not into an ever-growing prompt.
A bloated standing context file taxes every single turn. Trim weekly. If the agent is confidently wrong, suspect the context before the model.