Is "Codex gets slower in long sessions" a model problem or a context-budget problem?
Charles-SG · reddit · 2026-11-03
The author keeps seeing the same pattern in long coding sessions: two agents on comparable models, one stays flat while the other degrades — slower first token, longer turns, and a context window that becomes unusable for follow-through. His claim: we're comparing context management, not inference speed.
Three different things get called "slow":
- Slower first token (prompt length, cache miss, queueing)
- Longer per-turn thinking (more history to re-read)
- More turns to finish (the model lost something and is rediscovering it)
The expected decay path: history grows linearly → tool returns dump file contents and search results into the window → compression triggers → detail drops → the model re-reads what it lost → turns go up. Whoever offloads context earlier — subagents, external files, summaries — looks stable without being faster in any single turn. Prompt cache hit rate fits the same frame: same weights, different context policy, opposite experience.
The counterargument he takes seriously: stable may just mean it gives up on you sooner. Compression is forgetting; a long task quietly summarized into shape will look fast and drift late. If so, the metric should be final quality × total tokens, not felt speed.
He asks the community: are you seeing degradation, and which of the three? Does per-turn token logging make it obvious? Has anyone actually measured this, or is it all vibes?
More from coding & agent
- Model swap cut costs 59% and latency 75% — and he still blocked the release — Fun_Employment6042 · 2026-11-03
- Context drift broke a production LLM deployment by turn 10 — no jailbreak needed — GasProfessional5544 · 2026-11-03
- Obsidian dev open-sources Pele, a Mermaid-compatible renderer 100-500x faster than mermaid.js — Teknium · 2026-11-03
- Agent-built eval harness spawns subagents via codemode in one durable session — mitsuhiko · 2026-11-03
- SurfSense runs Claude Code and opencode as dual engines — lessons from a local NotebookLM — FurtiveMirth · 2026-11-03
- Impeccable crosses 100k GitHub stars, with weeks of releases teased ahead — jamdac · 2026-11-03