Codex and Claude Code advisor trees route cheap forks through Jev
A comparison shows Codex and Claude Code building agent trees where cheap subagents do the work and a strong model sits behind gates, with Jev routing cheap forks like file, tool and retry choices at a 302 ms median.
Codex and Claude Code take one flag and one prompt each and hand back an agent tree where the strongest model sits on call at 3 gates One flag, one rule, and the expensive model drops off the loop. Same pattern in both runtimes. The session model writes code, cheap agents scout and patch in parallel, the senior model sits behind a gate and only steps in when the session asks. → CODEX Sol on high writes code; explorer, worker, researcher at medium; Astra on call as architect, spawned at 3 gates → ADVISOR Sonnet 5.5 runs the session with three Haiku 5.5 subagents; Opus 5.5 on call, reads full transcript when consulted → ULTRACODE Opus writes a workflow script; Haiku scans, Sonnet patches in isolated copies; verify stage filters findings before Opus reads them → JEV routes the cheap forks — which file, which tool, retry or stop; 302 ms median per call (single-author replay) → GATE approvals_reviewer = auto_review; diff first, no edits until go set on the CLI: --advisor, --effort ultracode set in config: approvals_reviewer = auto_review, agents.default_subagent_model, model_reasoning_effort, gpt-6-luna, gpt-6.1-sol The model decides when to call the advisor, so write the three-call cap into CLAUDE.md or AGENTS.md yourself. Full tree and every flag mapped below ↓
This post quotes another post: x.com/mojunel/status/2111100246189089085