Episode 94 · Harness Design series
Context Is the Bottleneck (Harness Design, Part 2)
Ines finds that more than half her agent's failed runs never wrote a line: they overflowed the context window. With Gideon she works through why the history fills up, context rot, lossy versus lossless compaction, and the five tiers in Fan, Zhang et al. (arXiv:2609.20804). On SWE-Bench Verified, managing context was worth 35.7 points at a 32k window but only 2.7 at 128k, and almost all of that came from stopping overflow (78.7% of unmanaged runs overflowed at 32k, versus zero for managed runs). Eliding first and summarising later was the cheapest policy. Recoverable recall was rarely called and added no accuracy. The episode also covers window arithmetic, surprises in the tables, how compaction fails, keeping the plan outside the history, learned versus static policies, vendor features, sub-agents and the limits of the evidence.
Listen to the show on your preferred podcast platform.
Inside this episode
- Cold open: the runs that never wrote a line
- Why the bucket fills
- A bigger window is not a free fix
- Lossy and lossless: two families of fix
- The five tiers
- How the staged policy works
- The arithmetic of a window
- Finding one: management matters most when the window binds
- The overflow story
- Reading the tables: the surprises
- Finding two: cheap first, expensive later
- Finding three: the photos in the drawer
- What the summary says
- How compaction goes wrong
- Why the plan lives outside the history
- The wider toolkit: maps, memory and sub-agents
- Static rules and learned habits
- The vendors' versions
- What the paper does not tell you about context
- What Ines changed
Production credits & AI disclosure
Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.