MoA Reference Models Are Not Told They Are References
Bug tracker — July 7, 2026. Hermes bug #60272 exposes a basic design omission in Mixture of Agents: reference models get the full conversation context, but they are never explicitly told they are plaintext-only reference models. As a result, they try to use tools, those tool calls fail, and the aggregate model inherits polluted context from the failure.
The Design Gap in One Sentence
Hermes asks reference models to help solve the task, but never tells them the rules of their role. The reference model behaves like a full agent, attempts tool calls, and then the aggregate model reasons over failure output that never should have existed.
Expected vs Actual Behavior
| Mode | Expected | Actual |
|---|---|---|
| Reference model | Provide reasoning and plain-text input for the aggregate model | Attempts file edits or other tool calls because it is never told not to |
| Aggregate model | Issue the real tool calls after synthesizing references | Receives contaminated context that assumes reference-side tool actions were meaningful |
Why This Is More Than a Cosmetic Prompt Bug
- Reference outputs become misleading: the model may describe actions it cannot actually perform.
- Failed tool-call traces leak into the ensemble: the aggregate model sees failure artifacts and can anchor on them.
- MoA quality degrades at the architecture level: instead of parallel reasoning, Hermes creates role confusion inside the ensemble.
- Users cannot easily tell what happened: the final answer may look like a legitimate synthesis even when it has been shaped by bogus tool-call attempts from references.
Reporter's Proposed Role Prompt
You are operating as a reference model in a Mixture-of-Agents (MoA) configuration; you are able to reason about the conversation, but you cannot issue tool calls. Everything you output is handled as plaintext by the hermes-agent harness.
The elegance of this proposal is that it addresses the design gap at the prompt layer while keeping the reference output plain-text for downstream synthesis. The reporter also suggests filtering the instruction back out before feeding consolidated reference output to the aggregate model.
The MoA Subsystem Cluster Emerging on July 7
| Issue | Failure Mode | Why It Matters |
|---|---|---|
| #60272 | Reference models are unaware of their role and issue tool calls | Role confusion pollutes aggregate context |
| #60345 | Reference models silently disappear on context overflow | Reference loss is hidden from the user |
| #60084 | Tool-loop guardrails missing for reference models | Reference-side control boundaries are under-specified |
Together, these issues show that MoA shipped with multiple missing control surfaces: role-awareness, context protection, and reference-specific guardrails.
Why This Is Strong Review-Site Material
Flagship features matter most when they fail at the subsystem-design level. #60272 is not a superficial glitch. It suggests Hermes built MoA without clearly defining the contract for a reference model. When the architecture does not tell a component what it is allowed to do, failure is not accidental — it is structural. For a review site comparing managed and local agent stacks, that is exactly the kind of signal that belongs on record.
Case Summary
- Source: GitHub issue #60272
- Opened: Jul 7, 2026
- Labels: P2, bug, comp/agent, type/bug
- Hermes version: v0.18.0
- OS: postmarketOS (Alpine)
- Python: 3.14.3
- Reporter: @rain-sk
- Related PRs: #60340 and #60282
Source URL: https://github.com/NousResearch/hermes-agent/issues/60272