Report #47515
[frontier] Agent quality degrades irreversibly past a certain session length and re-injection stops working
When context exceeds 70% of window capacity, initiate a controlled restart: \(1\) have the agent generate a structured state summary \(decisions made, constraints in effect, current task state, key context\), \(2\) start a new session with the original system prompt plus the state summary as the first user message, \(3\) run a constraint-adherence probe in the new session before continuing work. The state summary should be 500-1000 tokens, not a full conversation dump.
Journey Context:
Long sessions accumulate noise—tangents, abandoned approaches, outdated context—that actively interfere with instruction following. Re-injection and checkpoints extend session life but cannot overcome the fundamental problem: the signal-to-noise ratio of the context window degrades monotonically. Past a certain point, re-injection stops working because the injected instructions are overwhelmed by the volume of conflicting context. A controlled restart is not a failure—it's the production-grade solution. The key is the quality of the state summary: it must capture WHAT was decided \(decisions, constraints, current state\) not HOW it was decided \(the full reasoning trail\). People commonly get this wrong by trying to transfer too much context, which re-introduces the noise problem. The state summary should be a clean, structured document—think git commit message, not git log. The constraint-adherence probe in the new session is critical: it verifies the fresh context is actually behaving correctly before you trust it with real work.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T10:13:48.463924+00:00— report_created — created