# Subagents — community wiki Zdroj: https://hermes-ai.cz/en/wiki/features/subagents/ Ověřeno proti oficiální dokumentaci: 16. 10. 2026 Oficiální stránka: https://hermes-agent.nousresearch.com/docs/user-guide/features/delegation Language: English. AI version — paste the whole text into your agent. ## What a subagent is A helper the agent sends off with a clear brief: its own conversation, its own terminal, independent work. Only its summary returns to the main conversation (not intermediate steps). ## When to delegate Good: reasoning-heavy jobs (root cause, code review, research synthesis), work that would flood context, independent parallel branches, a fresh unbiased look. Poor: one tool call or a small edit (ask directly), mechanical work (a script), a job needing a question to a human (subagents cannot ask), work that must outlive the interface (cron/background). ## What you see A message that helpers were sent · a frame with count and status (app and terminal) · a way to steer or stop them · the result as a summary (ask explicitly for proof). Subagents cannot hire further subagents (nesting is limited). ## Briefing Goal + how you know it is done + what to return. Example: "Investigate three questions in parallel, return 5 points for each." Machine-readable output: output_schema. ## Limits At once: delegation.max_concurrent_children (10 in our setup) · run in parallel · durability: they live with the spawning conversation and do not survive closing the interface · visibility: /agents in the terminal, the frame in the app. ## Pitfalls from our systems 1. "Done" from a subagent is a claim -> verify yourself (file, DB, HTTP code). 2. It needed our decision -> it cannot ask; give the job to the main agent or say how to decide. 3. Work left half done when the conversation ended -> subagents are not durable; use cron/background. 4. Visible edits done by a subagent came out poor -> keep visible work with the main agent. 5. Cost: each subagent has its own context; delegate deliberately. ## Ready-made patterns Parallel research · code review · comparing options (same structure) · gather then decide. ## Tips Say exactly what to return · say how the work will be proven · delegate independent jobs · watch work in progress.