Practice — Context Engineering for Agents (5 questions)

Your team's internal coding agent works well for the first ~30 minutes of a session and then noticeably degrades: it starts re-reading files it already read, forgets a constraint the user stated early on ("don't touch anything under /legacy"), and occasionally repeats a fix it already tried and confirmed didn't work. Session logs show it's making heavy use of a run_tests tool that returns the full stdout/stderr of the test runner verbatim on every call, and a read_file tool with no size limit.

1. Diagnose the mechanism behind the degradation — not just "the context got too long," but specifically why forgetting an early constraint and repeating a failed fix are the symptoms you'd expect from this cause.
2. Propose two concrete tool-design changes and explain, mechanistically, how each reduces the problem rather than just "using fewer tokens."
3. Independent of the tool fixes, what would you change about how the agent handles the /legacy constraint specifically, so it survives even if context does eventually get compacted?