How 12 teams assemble agent context: fixed prefixes for cache, retrieval policy set by miss cost

Gooner-Psy · reddit · 2026-10-26

Follow-up to the author's 12-team agent-harness comparison, this long piece covers one job: assembling context.

Consensus: keep the prompt start fixed

Instruction files stay small: OpenAI's harness team uses a 100-line AGENTS.md pointing elsewhere; Codex caps combined instruction files at 32 KiB; Stripe uses global rules sparingly

The split: who fetches the rest

Author's stance: cheap misses → let the agent search; customer-facing misses → retrieve for recall and pay for tokens. Gap: nobody publishes cost per successful task. The post doubles as a test case for his own pipeline, with a planted contradictory source.

Original post →

More from coding & agent

coding & agent channel →