Mixture of Agents

Mixture of Agents is a virtual model provider. Each named MoA preset appears as a selectable model under the moa provider.

When you select a MoA preset, the preset's aggregator is the acting model. It is the model that writes the assistant response and emits tool calls. Reference models run first and provide analysis for the aggregator to use.

Use MoA when a hard task benefits from multiple model perspectives but still needs Hermes' normal agent loop: tool calls, follow-up iterations, interrupts, transcript persistence, and the same session context as any other message.

You can select a preset through the normal model picker surfaces:

/model default --provider moa
/model review --provider moa

MoA presets are selectable on every Hermes surface, because MoA is a normal provider in the model system:

• CLI / gateway / TUI /model — /model <preset> --provider moa, or /model --provider moa for the default preset. A bare /model <preset> also works when the name exactly matches a configured preset.
• hermes model and the Dashboard model picker — a Mixture of Agents provider row appears with your preset names as its models.
• Desktop GUI app — the model dropdown shows an MoA presets section; selecting one (MoA: <preset>) switches the active model to that preset. The Desktop settings panel also creates and edits presets.

Configured presets therefore show up wherever you would pick any other model.

/moa is one-shot convenience sugar. It runs a single prompt through the default MoA preset, then restores whatever model you were on:

/moa design and implement a migration plan for this flaky test cluster

Hermes temporarily switches to the default MoA preset for that one turn, sends the prompt, then restores your previous model afterward. The whole argument is the prompt — /moa no longer interprets it as a preset name.

/moa

Bare /moa (no prompt) just prints usage.

To switch to a MoA preset for the rest of the session, select it from the model picker — MoA presets appear under a Mixture of Agents provider in every model-selection surface (see above). /moa is deliberately not a model switch, so a normal prompt can never accidentally change your model.

For each main model call when provider moa is selected, Hermes:

1. resolves the selected preset by name;
2. runs the configured reference models without tool schemas (they receive only the conversation's user/assistant text — not the Hermes system prompt or tool-call transcript — so reference calls stay cheap and avoid strict-provider rejections);
3. appends the reference outputs as private context for the aggregator;
4. calls the configured aggregator with the normal Hermes tool schema;
5. treats the aggregator response as the real model response;
6. if the aggregator calls tools, Hermes executes those tools normally;
7. on the next model iteration, the same MoA process runs again over the updated conversation, including tool results.

Because MoA is selected through the normal model system, it composes automatically with /goal, gateway sessions, TUI sessions, and Desktop chat.

You can configure named MoA presets from:

• Dashboard → Models → Model Settings → Mixture of Agents
• Desktop app → Settings → Model → Mixture of Agents
• hermes moa configure [name]
• config.yaml

The config stores explicit provider/model pairs, so you can mix providers and use multiple models from the same provider:

moa:
  default_preset: default
  presets:
    default:
      reference_models:
        - provider: openai-codex
          model: gpt-5.5
        - provider: openrouter
          model: deepseek/deepseek-v4-pro
      aggregator:
        provider: openrouter
        model: anthropic/claude-opus-4.8
      # Optional: pin sampling temperatures. When omitted (the default),
      # temperature is NOT sent and each model uses its provider default —
      # the same behavior as a single-model Hermes agent.
      # reference_temperature: 0.6
      # aggregator_temperature: 0.4
      max_tokens: 4096
      enabled: true

Default preset:

• reference: openai-codex:gpt-5.5
• reference: openrouter:deepseek/deepseek-v4-pro
• aggregator / acting model: openrouter:anthropic/claude-opus-4.8

Tuning advisor speed with referencemaxtokens

Each turn, MoA runs the reference models (advisors) in parallel and then the aggregator acts. Advisor generation is the dominant per-turn latency — turn wall time correlates strongly with how many tokens the advisors emit, because the turn waits for the slowest advisor to finish writing. By default advisors are uncapped (reference_max_tokens unset), so they may write long, essay-length advice.

Set reference_max_tokens on a preset to cap advisor output and give concise advice instead. The aggregator only needs the gist of each advisor's judgement, so a cap (e.g. 600) measurably cuts per-turn wall time with little quality impact. It caps advisors only — the acting aggregator's output (the user-visible answer) is never capped.

moa:
  presets:
    fast:
      reference_models:
        - provider: openrouter
          model: anthropic/claude-opus-4.8
        - provider: openrouter
          model: openai/gpt-5.5
      aggregator:
        provider: openrouter
        model: anthropic/claude-opus-4.8
      reference_max_tokens: 600   # concise advice → faster turns

Leave it unset (or 0/blank) to keep the prior uncapped behavior.

Advisor cadence with fanout

By default the advisors run once per user turn (fanout: user_turn) — they synthesize plan-level advice on the first message of the turn, then the acting aggregator works through the rest of the tool loop alone. This is the cheapest cadence: advisor cost does not multiply with the number of tool calls in a turn. Two alternative cadences trade cost for advice freshness:

• fanout: per_iteration — advisors re-run on every tool iteration, so their advice always tracks the latest tool results — at the cost of multiplying advisor latency and spend by the number of tool calls in a turn.
• fanout: every_n:3 — the middle ground: advisors run on the first iteration of each user turn and then every 3rd tool iteration (any N >= 2 works). Iterations in between reuse the cached guidance from the last advisor run, so the aggregator still gets advice on every step — it is just refreshed every N steps instead of every step. The counter resets on each new user message, so every turn starts with fresh advice. The mapping form fanout: {mode: every_n, n: 3} is also accepted and normalized to the string form.

moa:
  presets:
    fresh:
      reference_models:
        - provider: openrouter
          model: anthropic/claude-opus-4.8
      aggregator:
        provider: openrouter
        model: openai/gpt-5.5
      fanout: per_iteration   # advisors refresh on every tool iteration

Unknown or malformed values fall back to user_turn.

Privacy filter for advisor outputs

Advisor outputs can echo sensitive data from the conversation — emails, formatted phone numbers, API keys, JWTs — into the reference blocks shown in the UI, saved MoA traces, and the aggregator prompt. moa.privacy_filter (off by default) redacts those surfaces:

moa:
  privacy_filter: display   # or: full

• display — redacts user-visible surfaces only: the labelled reference blocks rendered in the UI and the records written by save_traces. The aggregator still receives the raw advisor text, so answer quality is unaffected.
• full — additionally redacts the advisor text injected into the aggregator prompt (and the one-shot /moa synthesis input).

Credential shapes (API-key prefixes, JWTs, private keys, DB connection strings) are masked by Hermes' central secret redactor; the MoA filter adds email and clearly formatted phone-number redaction on top. Patterns are deliberately conservative for code-review-style advice: bare digit runs, line numbers, timestamps, git SHAs, and IP addresses are never touched — only delimited phone formats like (555) 123-4567 or 555-123-4567 match.

Per-slot reasoning effort

Reference and aggregator slots may also set reasoning_effort. Use this when you want the same model to contribute at different depths, or when the aggregator should think harder than the advisory references. Valid values match Hermes' normal reasoning controls: none, minimal, low, medium, high, xhigh, max, and ultra.

moa:
  presets:
    deep_review:
      reference_models:
        - provider: openai-codex
          model: gpt-5.6-sol
          reasoning_effort: low
        - provider: openai-codex
          model: gpt-5.6-sol
          reasoning_effort: xhigh
        - provider: xai-oauth
          model: grok-4.5
      aggregator:
        provider: openai-codex
        model: gpt-5.6-sol
        reasoning_effort: high

Omit reasoning_effort to use the provider/Hermes default for that slot.

On HermesBench, a two-model MoA preset — claude-opus-4.8 aggregating over a gpt-5.5 reference — outscores either model run on its own:

The MoA configuration beats its strongest component (opus-4.8) by ~6 points, confirming that aggregating a second perspective lifts quality on hard tasks rather than just averaging the two.

MoA is built so the main conversation's prompt cache is never broken. Selecting a MoA preset is a normal model selection: it does not mutate past context, swap toolsets, or rebuild the system prompt mid-conversation. Your conversation history, system prompt, and tool schema stay byte-stable, so the cached prefix every other model relies on is preserved exactly as it would be for a plain model. Switching to or away from a MoA preset costs the same cache invalidation as any other /model switch — no more.

Both internal call types cache normally:

• Reference models receive a trimmed, deterministic view of the conversation (system prompt and tool transcript stripped — see the loop above). Because that view is a stable function of the stable history, a reference model's prompt prefix repeats across iterations and caches normally. References are short advisory calls with no tools.
• The aggregator is the acting model. The reference outputs are appended to the end of the latest user turn as private guidance. Because that text sits at the tail — below the entire stable prefix (system prompt + prior history) — it does not invalidate any cached prefix: the aggregator gets a cache hit on everything above the injection, and only the freshly appended tail is new. That is exactly how every normal turn behaves, where each new user message is also uncached tail tokens.

So MoA does not sacrifice prompt caching on either call type. Its only real cost is the extra reference calls per iteration — you pay for multiple model perspectives, not for broken caches. The long-lived conversation prefix shared with the rest of Hermes is fully intact.

• MoA is no longer listed under hermes tools; there is no moa toolset to enable.
• Setting enabled: false on a preset disables the reference fan-out for that preset: the aggregator acts alone, exactly as if you selected it as a plain model. This is the per-preset off switch surfaced in the dashboard and desktop settings.
• A preset's aggregator cannot be another MoA preset. Recursive MoA trees are intentionally blocked.
• Credential failures on one reference model do not abort the turn. Hermes includes the failure in the reference context and continues with whatever models returned.
• MoA increases model-call count. A single model iteration can involve multiple reference calls plus the aggregator call.