Jev handles mechanical routing in a layered Claude Code orchestration

winkle. @w1nklerrOriginally posted 2026.10.08

Open original on X

this is kind of a big deal: Anthropic just made frontier models less necessary put Opus on call. Let Sonnet build. Let Haiku swarm. claude --advisor opus --subagents haiku Sonnet 5.5 drives the primary session on high effort, writing diffs, implementing logic, and running test suites Haiku 5.5 subagents swarm the repo in parallel at 340 tok/s, handling file discovery, AST parsing, grep scans, and docs lookup Opus 5.5 stays in the background as the advisor, only waking up when the task crosses a real reasoning boundary → before the plan locks: does the architecture miss auth invariants or schema contracts? → when a test breaks twice: are we fixing the root cause or going down a recursive rabbit hole? → before calling done: did the final diff introduce hidden regressions or break pre-flight rules? Opus 5.5 reviews. Sonnet 5.5 builds. Haiku 5.5 swarms. JEV takes the same idea one layer deeper: mechanical decisions like which file to open, which tool to invoke, or whether to retry execute in roughly 16ms That means the expensive frontier models only wake up when execution paths actually diverge Plan on high. Delegate on medium. Keep Opus on call. - Opus 5.5 reads the full session history and catches deep architectural traps - Sonnet 5.5 drives edits, writes core logic, and executes the test harness - Haiku 5.5 subagents handle AST parsing, grep scans, file discovery, and API docs in parallel - JEV resolves 1,500+ mechanical routing branches in under 16ms - the advisor stays silent during routine execution to protect context paste this setup into Claude Code below: "Configure my Claude Code workspace for hierarchical advisor orchestration: 1. Audit ~/.claude/settings.json and project config for model roles: > Pin main session: sonnet with effortLevel: high > Pin subagents: haiku with effortLevel: medium > Pin advisor: opus for strategic review 2. Enable the advisor configuration: > Set advisorModel to claude-opus-5-5 > Enable a parallel subagent dispatcher pool with 3 workers 3. Configure automatic advisor checkpoints: > Consult /advisor opus before finalizing multi-file plans > Summon /advisor opus when the same test or compiler error fails twice > Run an advisor diff audit before declaring the task complete 4. Constrain subagent scopes: > Haiku returns structured AST summaries and docs snippets without modifying the main session Show every configuration diff first. Do not apply edits until confirmed." http://code.claude.com/docs/en/advisor

This post quotes another post: x.com/w1nklerr/status/2096964317107077567