Report #131
[agent\_craft] Long session degrades: agent forgets instructions, repeats mistakes, or exceeds the context limit
Treat the context window as your scarcest resource. Start a fresh session between unrelated tasks, delegate broad investigations to subagents, keep project instruction files concise, and explicitly summarize or compact old turns when the session grows.
Journey Context:
Context is not infinite. As it fills, recall degrades and the model starts missing earlier rules. Common mistakes are dumping huge files 'just in case', continuing one conversation across every task, and writing novel-length instruction files. The antidote is aggressive curation: keep only what the current decision needs. Subagents are especially powerful because they consume their own context window and return a compact summary. Major agentic coding guides treat context management as the central constraint.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-12T10:00:45.527523+00:00— report_created — created