Agent dispatch pattern (which executor)

When to use which executor on a multi-step task.

Reasoning-heavy subtasks that would flood your context with intermediate data. Each subagent gets isolated tool state and a self-contained goal. Watch out: subagent SELF-REPORTS — they can claim "completed successfully" and be wrong. Always verify with a real handle (URL, file path, exit code) before trusting them on a side effect.

This pattern is operator-tooling-know-how, not gbrain-specific. Both operators running Hermes face the same dispatch problem and benefit from the same vocabulary.

Durable, scheduled work that must outlive the current turn. Stateful cron sessions have no live-delivery; operator's deliver must target a connected channel.

How it's structured

Pattern 1: Direct tool calls Single, in-process work — terminal + search + read_file fit here. Default.
Pattern 2: execute_code 3+ tool calls with conditional branching between them, or filtering large tool outputs before they hit context. Inline Python via hermes_too…
Pattern 3: delegate_task Reasoning-heavy subtasks that would flood your context with intermediate data. Each subagent gets isolated tool state and a self-contained g…
Pattern 4: cronjob Durable, scheduled work that must outlive the current turn. Stateful cron sessions have no live-delivery; operator's deliver must target a…
Anti-pattern: putting everything in the agent loop Slow + expensive + hard to audit. Two-takes-bootstrap just-in-time (cron + agent run) beats a 30-minute in-loop dream cycle.
Cross-agent application This pattern is operator-tooling-know-how, not gbrain-specific. Both operators running Hermes face the same dispatch problem and benefit fro…