| name | hermes-budget-fields |
| description | When user asks about "iteration budget" or "turn limit" for Hermes or AO worker, use this skill FIRST. It documents the 3 distinct fields (agent.max_turns, delegation.max_iterations, and AO's NO-iteration-concept) so you never conflate them. Triggers on phrases like "iteration budget", "turn limit", "max_iterations", "how many iterations", "agent loop", "change to 100/300/500/1000", "what's the cap on". |
Hermes iteration / turn budget fields
The 3 fields (and which is which)
When the user asks about "iteration budget" or "turn limit", they may mean any of these:
| # | Field | Where | Default | What it actually caps |
|---|
| 1 | agent.max_turns | ~/.hermes_prod/config.yaml line 20 | 1000 (as of 2026-06-27) | Hermes chat-agent loop (the main agent in a Slack/Cron session). One turn = one LLM call + tool executions. |
| 2 | delegation.max_iterations | ~/.hermes_prod/config.yaml line 348 | 500 | Subagent (delegate_task) recursion depth. How deep a subagent can recurse before the parent is forced to stop. |
| 3 | (none) | n/a | n/a | AO worker has NO iteration budget. The loop is the model's tool-call loop, terminated by child_timeout_seconds: 600 (10 min) or max_spawn_depth: 1 (no recursion past one level). |
Bug-ref 2026-06-27 02:32 PT: I conflated all three in a /roadmap report. User corrected: Hermes max_turns is 1000 (not 300); AO worker has no iteration concept (verified by grep, zero iter matches in agent-orchestrator.yaml). mem0 entries saved.
How to verify (the 30-second check)
grep -n "agent:" $HOME/.hermes_prod/config.yaml | head -3
grep -n "max_turns:" $HOME/.hermes_prod/config.yaml | head -3
grep -n "delegation:" $HOME/.hermes_prod/config.yaml | head -3
grep -n "max_iterations:" $HOME/.hermes_prod/config.yaml | head -3
grep -inE "iter|iteration" $HOME/.openclaw/agent-orchestrator.yaml
If step 3 prints anything, the config has drifted. AO workers do not count iterations — they count tool calls and wall-clock time.
How to change
| User asks | File | Field | Line |
|---|
| "Change Hermes chat budget" | ~/.hermes_prod/config.yaml | agent.max_turns | 20 |
| "Change subagent recursion cap" | ~/.hermes_prod/config.yaml | delegation.max_iterations | 348 |
| "Change AO worker time/count" | ~/.openclaw/agent-orchestrator.yaml | child_timeout_seconds / max_concurrent_children / max_spawn_depth | (no iter field) |
Always run the verify above BEFORE editing. The numbers in the table above are correct as of 2026-06-27 — confirm they're still current before assuming.
Common confusions
- "iteration budget 300" → user may mean field 1, field 2, or both. Ask which, or grep both and report current values to user before changing.
- "AO worker iteration cap" → there is none. Point user at
child_timeout_seconds: 600 instead.
- "max_iterations 100" → that's field 2 (delegation), not AO. Don't change
agent-orchestrator.yaml thinking it controls this.
- "Hermes crashes after N iterations" → field 1 (
agent.max_turns). Crash usually means the LLM hit max_turns and the gateway returned an error rather than continuing.
- "Subagent hangs" → field 2 OR
child_timeout_seconds. Run the verify and check both.
Why this matters
The user has asked about iteration budget 4 times in 7 days across different threads (C0AJ3SD5C79 / 1782436438, C0AJ3SD5C79 / 1782379496, etc.). Each time I had to re-derive which field was being asked about. This skill makes the answer 30 seconds instead of 30 minutes.
Related
agento SKILL.md § "Per-project concurrent-spawn lock" — also talks about max_concurrent_children but not as iteration budgetdelegation.max_concurrent_children: 3 — caps how many subagents run in parallel, NOT iterationsdelegation.max_spawn_depth: 1 — caps nesting depth, NOT iterations- mem0 entry
hermes-budget-fields-three and ao-no-iteration-concept
