saved
The Oracle and the Firm
gist
Calvin French-Owen splits frontier long-horizon work into two context strategies. OpenAI's Codex is an "oracle": one long thread, server-side compaction in the responses API, coherent details, serial tokens. Anthropic's Claude Code is a "firm": Explore and review sub-agents on clean windows that return summaries, which looks faster and busier but duplicates work, costs more, and drops facts a child never reported up. He expects the labs to steal from each other—less-lossy compaction at Anthropic, multi-agent training at OpenAI.
ideas
- Long work is a context-management problem. Frontier models have to explore via tools and thinking across huge token counts; scaling that window is how harder, longer tasks get solved.
- Codex is an oracle thread. Server-side compaction in the responses API keeps one coherent trajectory, preserves small details, and can change without client upgrades while improving K/V cache routing.
- Claude Code is a firm. It splits the problem across Explore/review sub-agents with their own windows and returns only summaries, like a human org that never sees each other's hidden state.
- The firm looks busy and forgets. Parallel tokens feel faster and cost more because of duplicate search; facts a sub-agent omits from the parent report go missing even after the research happened.
- The end state mixes both. Anthropic's compaction is too lossy and client-bound; OpenAI still under-trains multi-agent setups. He expects each lab to adopt the other's strength.
quotes
“For any frontier model, accomplishing real work is an exercise in context management.”
“Codex typically keeps one long thread going with frequent compactions.”
“In effect, this looks more like the way human organizations run.”
“If a sub-agent deemed a fact was not worth reporting back to the parent agent, then it will be missing from the context.”