Files
agent-estate-wiki/concepts/tool-calling.md
T

2.4 KiB
Raw Blame History

title, type, status, created, updated, verified_on, confidence, tags, sources
title type status created updated verified_on confidence tags sources
Tool Calling concept active 2026-07-22 2026-10-20 2026-10-20 high
concept
tool-calling
raw/configs/hermes-config-sanitized.txt

Tool Calling

Idea

How Hermes decides to invoke a tool, and — more specifically for this page — how it detects and stops itself when tool use goes wrong (loops, repeated failures).

How this install implements guardrails

Verified (~/.hermes/config.yaml, tool_loop_guardrails:):

warnings_enabled: true
hard_stop_enabled: false
warn_after:
  exact_failure: 2
  same_tool_failure: 3
  idempotent_no_progress: 2
hard_stop_after:
  exact_failure: 5
  same_tool_failure: 8
  idempotent_no_progress: 5
  • Warnings trigger well before any hard stop (2–3 repeats vs 5–8).
  • hard_stop_enabled: false — Hermes will warn about a stuck tool loop but will not currently force-stop itself. That's a deliberate-looking but consequential setting: a genuinely stuck loop only gets a warning, not an automatic halt.

Why it matters here

This is the mechanism behind runbooks/troubleshoot-tool-loop — that runbook's "agent loops on one tool" symptom is exactly what warn_after/hard_stop_after are meant to catch. With hard-stop disabled, a human (or the messaging integrations) is the actual backstop today.

Re-verified 2026-10-20 — the tool_loop_guardrails: block is byte-for-byte unchanged. Live values in ~/.hermes/config.yaml: warnings_enabled: true, hard_stop_enabled: false, warn_after: {exact_failure: 2, same_tool_failure: 3, idempotent_no_progress: 2}, hard_stop_after: {exact_failure: 5, same_tool_failure: 8, idempotent_no_progress: 5}. 90 days of no drift on this guardrail.

The 2026-10-18 incident is a different loop and the guardrails did not catch it. The degenerate-stream case (see incidents/2026-10-18-degenerate-stream-canned-substitute) was a model output repetition loop — the model repeating itself in the response stream, not a tool call repeating. Hermes's degenerate-stream guard cut it and delivered a canned non-answer, but tool_loop_guardrails (which watches tool calls) never fired. Two separate loop-detection mechanisms exist; do not assume one covers the other.