Agent Dispatch Sequence

The execution-mode pattern. Opus stays in main and dispatches mechanical work to a Sonnet subagent with the 8-element prompt contract. Then Opus independently verifies the result before declaring complete.

When dispatch fires

Per G-MODEL-ROUTING-AT-EXEC, Opus dispatches Sonnet whenever execution mode is triggered. Triggers include:

• /gsd:* execution commands
• Execution skills (/acquisitions-outreach, /dispo-blast, /hubspot-deal-ingest, etc.)
• TodoWrite task transitions to in_progress for mechanical work
• Plan file checkbox transitions in ~/.claude/plans/*.md
• User keywords: “implement”, “execute”, “build it”, “ship it”, “go”, “do all of them”
• Multi-file batch edits against an already-defined rubric

Trust-but-verify

Sonnet’s report describes what it intended to do. Opus runs the independent acceptance check before marking complete:

• Code edit? → grep for the new content; verify old content removed
• File created? → check existence + size
• Tests added? → count test definitions; run the test suite
• Build passed? → re-run the build with same args; verify exit 0

If verification fails, Opus either re-dispatches with corrected prompt or does the work directly.

Why this pattern

Three reasons:

1. Cost — Sonnet is ~5× cheaper per token than Opus. Mechanical work doesn’t need Opus brain.
2. Context isolation — subagents run in fresh contexts. They don’t pollute Opus’s working memory with intermediate file reads.
3. Parallelism — Opus can dispatch many Sonnets concurrently for independent work (e.g., 5 parallel content authoring tasks).

See Model Routing for the full decision tree.