Subagent Commander

Finding

Complex Hermes work often collapses research, analysis, QA, and recommendation into one oversized agent run, which weakens evidence quality and makes errors harder to catch.

Current

A typical Hermes installation can delegate work, but delegation is often used informally: one helper researches, another reviews, and the main agent summarizes without a stable operating pattern. The weak point is not the tool itself; it is the lack of a repeatable command structure for when to split work, which roles to use, what each role may change, and what evidence must come back before the coordinator makes a decision.

Suggested

Define a standard Scout / Analyst / QA delegation pattern for broad or high-stakes tasks. Exact change: add a short Subagent Commander section to SOUL.md or the main operator runbook stating: “Use Scout for source collection, Analyst for synthesis, QA for critique; the coordinator owns the final recommendation and must verify key claims before acting.”
Add role-specific prompt templates for delegated work. Exact change: create or update a skill such as subagent-research-pattern with three reusable prompt blocks: Scout: collect sources and cite URLs/files, Analyst: compare evidence and identify options, and QA: find gaps, risks, hallucinations, and missing verification.
Add a verification habit before accepting delegated summaries. Exact change: update the coordinator checklist, cron prompt, or task runbook with: “Before final response, verify at least one cited source or file path from each subagent, reject unsupported conclusions, and clearly separate evidence from recommendation.”

Impact

This makes Hermes delegation operational instead of decorative. Scouts reduce search bias, Analysts keep synthesis focused, and QA catches weak assumptions before they reach the user or a public page. The coordinator stays accountable, so delegated work improves output quality without creating uncontrolled autonomous side effects.

Effort

Small — it requires one operating rule, one reusable skill or prompt template, and one verification checklist item rather than new infrastructure.

Public page note

Safe public content includes the maturity principle, generic Scout / Analyst / QA roles, delegation boundaries, and verification habits. Internal-only content includes private prompts, raw subagent outputs, session transcripts, customer data, credentials, local file contents, and any operational details that reveal sensitive workflows.