Skip to content

ADR-0018: Hermes Agent is not adopted in any role; Hermes models remain one candidate for a future self-hosted AU inference slot

  • Status: Accepted
  • Date: 2026-07-31
  • Deciders: El
  • Related: ADR-0001 (orchestration-first), ADR-0009 (LLM provider strategy — self-hosting on AU infra left open), the llm-client provider-abstraction boundary

Context

Two different things ship under the name "Hermes" (Nous Research): the Hermes models (open-weight LLMs) and Hermes Agent (an open-source, MIT-licensed autonomous agent runtime — a self-hosted daemon with persistent memory, self-writing skills, cron, messaging gateways, sandboxed execution and subagent delegation, released Feb 2026). We evaluated whether Hermes Agent could serve as, or accelerate, the ABOS platform — as the engine, as a per-client single-tenant deployment, or as a leaf-node executor behind our gate.

Decision

Hermes Agent is not adopted in any role — not as the BOS, not as the engine, not as a per-client managed offering, not as a leaf-node executor.

The reasoning is structural, not a knock on its quality:

  1. It is single-tenant by design ("the agent that grows with you"): one box, one user, one accreting memory. Adopting it fights our multi-tenant, per-client-isolated model rather than composing with it. Serving N isolated clients on Hermes means N instances plus building provisioning, the registry control plane, per-tenant secrets, billing, a cross-tenant operator view and exit anyway — i.e. the hard 70% is ours regardless, with Hermes reduced to a leaf.
  2. Its default posture inverts our governance. Its defining feature is self-writing, self-evolving skills and autonomous terminal/file access — a large blast radius. ABOS is gate-first ("No review, no go"). A gate bolted onto a runtime designed to route around one is a leash on a foreign, fast-moving base (≈fortnightly upstream releases we do not control), re-verified forever by two people. Once the gate, isolation, audit, tenancy and exit are built and self-evolution is disabled to make it safe, the part of Hermes that had value is the part we neutered; what remains is commodity plumbing.
  3. Provider-agnosticism is already ours via llm-client; Hermes' model-switching adds nothing.

Hermes models (Hermes 4 70B/405B, 4.3 36B — open weights) remain one candidate — alongside Llama, Qwen and Mistral — for the self-hosted-open-weights-on-AU-infrastructure slot behind llm-client, exercising the door ADR-0009 left open. No commitment now; chosen at provisioning time on eval against our tasks, same discipline as VPS-provider-agnosticism (ADR-0014 topology is a swappable implementation detail). Do not hardcode "Hermes" into docs any more than "Hetzner".

Consequences

We continue building the differentiated layer (multi-tenancy, governance, recipes, learned knowledge, exit) ourselves, and borrow commodity plumbing (messaging ingest, sandbox abstraction, memory tiering) as patterns or components rather than adopting an opinionated runtime. Hermes remains useful as a reference implementation and competitive datapoint; its own community's "treat auto-generated skills as drafts needing review" consensus is external validation of our gate-first stance, usable when justifying governance overhead.

Alternatives considered

  • Hermes Agent as the BOS / engine — rejected: single-tenant shape, inverted governance, loss of IP/moat (MIT engine was never ours to retain), roadmap becomes downstream of Nous Research.
  • Reselling per-client managed Hermes deployments — rejected: bespoke-consultancy economics (thin margin, ops-heavy per-client snowflakes) that compete with the templated-product thesis.
  • A hardened, self-evolution-off Hermes as a leaf-node executor — rejected: for the bounded, gated tasks ABOS runs, n8n + the LangGraph brain already execute them inside our isolation and gate; Hermes brings no capability gain while adding a second heavyweight runtime to harden, patch and monitor, violating engine-agnostic discipline.

Amendment (2026-07-31): scope clarification — consulting delivery

This ADR governs the ABOS platform's own architecture: ABOS is not built on, and does not adopt, Hermes Agent. It does NOT preclude delivering Hermes-based solutions to a client as a paid consulting engagement — that is a separate commercial decision, unaffected by this ADR. The rejection is "we don't build our platform on Hermes," not "we never do Hermes work for clients."