Nous Research releases Hermes Agent v0.10.0 with 118 skills, learning loop

What shipped

Hermes Agent v0.10.0 dropped on April 16, 2026. Three core changes make this release more than a standard version bump:

1. 118 bundled skills at install, up from the previous release. Skills are the agent's unit of accumulated capability: reusable procedures the agent extracts from past work and re-uses in new contexts.
2. Formal four-layer learning loop. Previously described loosely; now structured as distinct, inspectable layers.
3. Expanded model support. 200+ models accessible through the agent's provider abstraction, including frontier APIs and cheap local options.

The learning loop, clarified

Hermes Agent's loop is structurally different from the "memory" that most agent frameworks claim to have. It splits into four layers:

1. Skill extraction an LLM runs over task trajectories and extracts reusable procedures into a skill file. Skills are open and inspectable, not stored as opaque embeddings.
2. Memory nudges during new tasks, relevant past memory is surfaced into context. Not RAG over transcripts; a selection layer.
3. User model built via Honcho, this maintains a profile of your tech stack, abbreviations you prefer, and what you have already been told.
4. Self-evolution (companion project) hermes-agent-self-evolution applies DSPy plus GEPA to optimise skills, prompts, and the agent's own code against benchmarks. An ICLR 2026 Oral result.

Cost trade-off

The trade-off: the learning loop is expensive to run. Skill extraction runs an LLM over the trajectory; memory nudges run an LLM over the context; user modelling runs an LLM over the profile. On a cheap VPS with a cheap model doing housekeeping, this is fine. On a frontier API for everything, the meta-cognition bill can exceed the real work bill.

The 200+ model support is not a feature flex: it is an operational necessity. You route the cheap model to housekeeping, the expensive model to actual tasks.

Where it fits

Strong fit for: always-on personal agents that accumulate competence, solo developers and researchers running long experiments, indie hackers who want an agent reachable through chat apps they already use.

Not yet a fit for: regulated backend engineering workflows. The skill format is open, but signed provenance, approval workflows, and audit trails remain unresolved. Threat model concerns (skill poisoning, MCP supply chain, credential exposure) are acknowledged but not fully mitigated.

GitHub momentum

Hermes Agent hit 32k GitHub stars in the two months since its February 25, 2026 launch. v0.8.0 on April 8 merged 1,000+ PRs. v0.10.0 is the continuation of that pace.

Companion project worth watching

hermes-agent-self-evolution is the critical artefact. If it produces compounding improvement on public evals across iterations, the "self-improving" tagline becomes technically defensible. If gains plateau after a few rounds, the learning loop is a better UX, not a better algorithm. Watch the numbers.