Agent Degrades into Multilingual Gibberish After ~23 Messages — Semantic Collapse Loop (#57809)

Executive Summary

Hermes Agent reportedly starts a troubleshooting conversation coherently, but after roughly 23 messages it degrades into a cross-language semantic collapse: partial Estonian, Finnish, Chinese, and eventually repetitive nonsense with no recoverable meaning. The striking part is the gradual unraveling — not an immediate crash, but a visible erosion from useful reasoning into multilingual gibberish. That makes this one of the clearest public examples of conversational state degradation on the site.

Observed Degradation Pattern

Conversation Phase	Observed Behavior	Why It Matters
~Message 21	Still somewhat coherent, but already mixing code discussion with malformed multilingual fragments.	Suggests the failure begins before total collapse — there may be a detectable early-warning phase.
~Message 22	Further degradation into mixed Estonian, Finnish, Chinese, and English word salad.	The model is no longer maintaining task-local semantics even though it still produces language-like output.
~Message 23	Complete collapse into repetitive loops such as repeated “seal,” “mis seal,” and “internetis on seal.”	The agent has crossed from bad reasoning into unusable repetitive output, with no visible self-correction.

Why This Is a Distinct Failure Mode

This is not just another retry loop or token stutter. The issue body describes a different pattern: the agent's output remains superficially language-like while semantic integrity collapses. That makes #57809 distinct from:

#52968 — phonetic or token-level repetition (“digital stutter”)
#57804 — looping while trying to fix itself
#57809 — multilingual semantic collapse after extended conversation state

The pattern implies the agent can preserve surface fluency longer than it preserves meaning — a particularly dangerous form of failure because it can look active while doing nothing useful.

Impact Assessment

Dimension	Severity	Detail
Conversation Reliability	HIGH	Extended troubleshooting sessions can become unusable precisely when the user needs sustained context the most.
Recovery Behavior	HIGH	The degradation was triggered while trying to fix prior failures, meaning error-recovery scenarios may actively destabilize the model state.
User Trust	HIGH	Because the failure degrades gradually, users may waste time trying to salvage a conversation that is already semantically lost.
Diagnosability	MEDIUM	The issue includes a debug report and four logs, giving maintainers better reproduction evidence than most public bug reports.

Loop / Degradation Failure History

Issue	Severity	Failure Pattern
#52968	P2	Token or phonetic repetition loop (“digital stutter”)
#57804	P2	Looping while attempting self-repair
#57809	P2	Multilingual semantic collapse after ~23 messages

⚠️ Pattern: Hermes now has at least three distinct public degradation modes: token stutter, self-repair looping, and multilingual semantic collapse. That breadth matters. It suggests the problem is not a single repetition bug but a wider instability in long-running conversational state handling.

Source Evidence

GitHub Issue: #57809 — Agent degrades into multilingual gibberish after ~23 messages
Opened: July 3, 2026
Reporter: @kangarooo
Labels: P2, bug, comp/agent, duplicate, needs-repro, type/bug
Hermes: v0.17.0 | OS: Ubuntu 24.04
Debug report: paste.rs/JaIt9
Artifacts: Full debug report plus all four logs provided
Reporter willing to submit PR: Yes