AI For Dev

nous-research

4 min read

Nous Research

Open-source AI research collective best known in the wiki for Hermes Agent — a self-improving open-source agent framework whose closed learning loop (GEPA) is one of the wiki’s most distinctive technical contributions to the agent-architecture thread. Nous Research occupies a niche between academic AI labs (single-paper-per-year cadence) and frontier commercial labs (Anthropic, OpenAI) — they ship working open-source agent systems with research-grade ideas baked in.

  • Founded: 2023 (community / collective origin)
  • Corporate structure: open-source research collective (not a traditional incorporated entity in the wiki’s product context)
  • Sites: nousresearch.com, hermes-agent.nousresearch.com
  • GitHub: github.com/NousResearch

Products tracked in this wiki

  • Hermes Agent — open-source self-improving agent framework. Positioned as an alternative to OpenClaw but distinguished by GEPA (“backpropagation for prompts, not weights”): every ~15 tool calls, the agent pauses, reviews failures, and updates its own prompts and skills without fine-tuning. ~12,000 GitHub stars within days of launch (March 2026 per Berman; revisited April 2026 per WorldofAI). Multi-platform gateway (WhatsApp/Telegram/Slack/TUI), Docker-based, supports local models via Ollama (Gemma 4 explicitly recommended).

Technical contributions tracked in this wiki

  • GEPA (closed learning loop) — see hermes-agent for the full description. The conceptual framing — “backprop for prompts instead of model weights” — is one of the cleanest explanations the wiki has for how an agent gets better without fine-tuning. Conceptual root traces to Karpathy’s AutoResearch work, but Nous’s GEPA is the most concrete production implementation.
  • Autonomous skill creation — when Hermes solves something, fixes an error, or is told to remember a task, it converts the experience into a reusable skill and pulls it later when relevant. Adjacent in spirit to Anthropic’s Claude Code skills ecosystem, but the creation loop is automatic.

In the wiki’s larger threads

  • Self-improving agents — Nous Research is the wiki’s anchor entry for “production-grade self-improving agent loops.” Hermes sits alongside ThePopeBot (Docker + GitHub Actions, no GEPA) and loop + OpenBrain (Anthropic-native primitives, no autonomous improvement) as the three architectural alternatives in the OpenClaw-replacement category.
  • Open-source agent frameworks — together with ThePopeBot, Paperclip, DeepCode, and the GStack / Superpowers / Agency / Hermes cluster of Claude Code skills, Nous’s work represents the open-source counterweight to commercial agent platforms.
  • Local-AI integration — Hermes Agent’s explicit recommendation of Gemma 4 via Ollama for fully-local operation reinforces the wiki’s local-AI cost-reduction thread.

How Nous Research differs from the frontier labs

Nous ResearchAnthropic / OpenAI / Google
FormOpen-source collectiveCommercial frontier lab
ReleasesOpen-source software (full agent frameworks)Proprietary models + APIs (mostly)
FundingCommunity / grants / openVC / corporate
Audit-abilityFull source visibleClosed model weights
Self-improvement IPOpen (GEPA is documented)Closed
Wiki-tracked policy positionsNoneOpenAI’s Industrial Policy paper

Nous Research is the wiki’s clearest example of the open agent-research alternative to commercial frontier labs. They ship research-quality ideas as working software rather than as papers, and the software is fully auditable.

Editorial framing the wiki applies to Nous Research sources

  • Nous’s own technical claims about Hermes Agent (GEPA, autonomous skill creation, self-improvement) are documented in their public docs at hermes-agent.nousresearch.com — treat as primary source for how Hermes works, but flag that effectiveness claims (“gets better the more you use it”) don’t have independent third-party benchmarks yet.
  • Coverage from WorldofAI and Matthew Berman is enthusiastic and demo-focused — discount the “this is revolutionary” framing but trust the demonstrated mechanics (install, skill management, the Obsidian-vault demo).
  • Watch for: independent benchmarks of GEPA (does it actually improve agent quality over time? by how much?) and follow-on releases from Nous beyond Hermes.

Open questions

  • Funding model — how is Nous Research sustainably funded? Community? Grants? Corporate sponsorship?
  • Beyond Hermes — what else is Nous Research working on? The wiki only tracks Hermes; their broader research output isn’t covered.
  • GEPA effectiveness — no independent benchmarks yet. The conceptual framing is sound but the empirical evidence is anecdotal.
  • Relationship to other open agent frameworks — does Nous collaborate with the Paperclip / GStack / Hermes-skills communities, or is each project independent?

See Also

Graph View

Backlinks