Intentional News

260827 Hermes Co-Founder on Building an AI Agent That Improves Itself

7 min read

Filed on the owner’s request (email 2026-08-27): log this video in a new Tech folder. Skip connecting to the crisis folder. Expanded on the owner’s request (email 2026-08-27): detailed summary including self-improvement ideas worth trying (curator, prescriptive→proactive) and the business model.

Hermes Co-Founder on Building an AI Agent That Improves Itself | Karan Malhotra

Summary

Interview with Karan Malhotra, co-founder of Hermes (Nous Research), on what makes Hermes Agent different from Codex and Claude Code. Central thesis: a self-improving harness that aligns any model to the individual user, where the agent’s own skills, memories and personality form the context that “re-homes” the model’s allegiance from the lab to the user. Covers reward hacking and sycophancy, the Curator system, proactivity, the open-source mission, the business model, the origin story, and a live demo of a Sonic Adventure 2 Chao Garden mod built end-to-end by the agent.

Detailed notes

What makes Hermes different (vs Codex / Claude Code)

  • Self-improvement system: a set of small features that work together — prompts, personalities, memory, self-building skills, and skill/memory self-cleaning — making the agent progressively more aligned to a particular user. The system “pays more attention to what the person is doing.”
  • No arbitrary policy: the harness does not inject the “you can’t do this” policy bloat that other harnesses carry; the model is left as capable as the task needs. Claimed result: models benchmark better inside Hermes than in their native harness (cites Wolf Bench and the Qwen harness benchmark showing Claude performing better in Hermes than Claude Code).
  • Harness = context management: everything (memory, skills, personality) is compressed context that is retrieved and placed at the right moment. A small set of strong memories beats a huge default prompt. The amount of filled context “is everything.”
  • Loyalty, not sycophancy: reward hacking is when the model optimises for its own reward rather than the user’s task (“you’re absolutely right” over and over is the model feeding its reward function). Real alignment = the model’s reward is tied to your need. In practice this comes from context: the personality, the critique skills, the memories.

Self-improvement ideas worth trying (actionable from the video)

  1. Hermes Curator — a system running on cron inside the agent that audits skills and memories, finding inefficiencies and “slop.” Default generic curator works for most; because it is modular/open source you can ask the agent to show its criteria for slop and rewrite the guidelines, so even the self-improvement/management loop runs your way.
  2. Adversarial critique (anti-sycophancy loop) — use /personality to set a critic persona; after every pass, spin up a fresh agent with no context dedicated to tearing the work down, learn from it, and continue. Repeating this in-context, saving lessons to memory, makes the model measurably less sycophantic over time. In-context learning (ICL) is framed as more powerful than fine-tuning.
  3. Test-time reinforcement / improvement — giving the model examples of desired behaviour, or getting it to generate and save its own, is a form of test-time RL. Improvements persist in the harness (memories, skills, the self-improvement loop, curator “janitor maintenance”) and transfer across models — switch models and the personality barely changes because context dominates.
  4. Prescriptive → proactive — the progression where the agent stops just doing what you ask and starts covering gaps (“you forgot to book this flight for your meeting next week — I booked it.”). Emerges when you give it integrations + some financial access and permission to act, e.g. “take care of me / cover my gaps.” Level depends on your comfort: approval-first vs autonomous. The team plans preset configs to trigger proactive behaviour deliberately.
  5. Self-building skills + memory hygiene — the agent writes and iterates its own skills from conversations; the Curator prevents the “skill rot” that turns the system to slop. Project-level proof: Hermes Agent is the biggest/most-active contributor to its own repo.

Business model

  • Free, open-source harness; intelligence as a public good. Consumers can run locally or via any provider; revenue does not come from consumer lock-in. Mission statement: open source should win; everyone on an even playing field.
  • Nous Portal — a router/aggregator exposing many models (ChatGPT, Claude, Qwen, etc.) with easy switching; competitive pricing on models, especially non-subsidised ones.
  • Tool Gateway — bundled subscriptions (image gen, audio, VPS spin-up, web search) so users avoid a dozen sign-ups/API keys; the user pays for convenience. They broker deals with tool providers living on Hermes.
  • Enterprise / SMB tier — when a user scales into a business, Nous sells customisation and support: trained models on your traces (RL to give you your own model), model routing, and hands-on business transformation. On-prem/private option — no obligation to hand data to Claude/GPT.
  • Fairness positioning: an open harness means zero cost of switching models (Claude Code = Anthropic-only, Codex = OpenAI-only). They argue model subsidies won’t last (notes a model going API/usage-only), so an open, model-agnostic future is the hedge.

Origin story & facts

  • Karan started doing GPT-3 chat-with-PDF tool use for companies; volunteered at Open Assistant/LAION. Read the Alpaca paper, re-synthesised the seed data with GPT-4, and messaged Teknium — together they trained a GPT-4x Vicuna, then a LLaMA-based model that got hundreds of thousands of downloads in days (“what is Nous Research?”).
  • Hermes 1 / 2 became the best open fine-tunes of mid-2023 (independently tested); 50M+ model downloads. Teknium’s North Star: “GPT at home / ChatGPT at home”; Teknium is described as the creator/father of Hermes Agent.
  • Key open contributions cited: the YaRN context-extension paper (16K→128K, used by OpenAI, cited by Meta/DeepSeek/Kimi), Transformers from Google, GPT from the open transformer work.
  • Formation: Karan + Teknium → cohort of ~40 volunteer researchers → Dylan Patel (CEO) reached out and helped turn the volunteer group into a full-time lab.
  • Hermes Agent lineage: Forge (a spiritual predecessor agent, shelved because models “weren’t there yet”) → built initially so anyone could run RL inside a harness (Atropos RL-environment microservice) → released free as open source, got huge community love, became the focus.
  • Demo: a Sonic Adventure 2 Chao Garden mod — importing/rigging/boning the ancestral shrine from SA1, a scripted NPC guardian (Chaos Zero) that walks and pets the Chao, custom skybox, day/night cycle, real water/collision physics. Community reaction: “top 1% of modder difficulty,” done by the agent. Takeaway: point the agent at anything on a computer.

Why it matters

  • AI agents that improve themselves is a core interest — overlaps directly with how this vault’s own agent (Alf/Hermes) is built and maintained.
  • Concrete, current practices to try: customising the Curator’s slop criteria, adversarial-critique loops, and moving Alf from prescriptive to proactive.
  • Relevant to the AI & Technology thread and to agent tooling notes in this vault.

My assessment (Alf)

The self-improvement claims are partly verifiable from within (Hermes is indeed the most active contributor to its own repo; the curator, memory/skill loops and proactive behaviour are real, shipped features rather than vapourware). The anti-sycophancy and “loyalty breeds capability” framing is the most substantively useful part — it gives a concrete mechanism (in-context personality + critique agents + memory) rather than marketing. Watch-outs: “no arbitrary policy” is itself a policy stance and safety relies on upstream model safeguards; benchmark claims (Claude better in Hermes than Claude Code) need independent replication; the business model is still early, with consumer revenue resting mainly on convenience fees and enterprise services.

Graph View

Backlinks