Nous Research Publishes Hermes Agent Operator's Manual, Detailing Self-Improving Framework Architecture

Nous Research’s Hermes Agent now has a public operator’s manual. Published on May 24 by independent technical writer Flaviu Vlaicu, the 30,000-word reference document covers installation, architecture, failure modes, and advanced configuration for the open-source agent framework that shipped its first public release on February 25, 2026. The manual, verified against Nous Research’s official documentation and the project’s GitHub repository, is the most comprehensive third-party technical reference for any agent framework published this year.

What the Manual Reveals

The document lays out Hermes Agent’s core architecture: a ReAct-style synchronous loop with six execution backends (local, Docker, SSH, Daytona, Singularity, Modal), provider-agnostic model routing, and a per-task turn cap of 90 by default to prevent runaway token consumption. These are known features. What the manual makes concrete is how they fit together in production.

Three systems distinguish Hermes from other agent frameworks:

GEPA (Genetic-Pareto Prompt Evolution) is an offline optimization pipeline for agent prompts. Rather than hand-tuning system prompts, Hermes generates candidate prompt variations, evaluates them against trajectory-based training data, and selects winners using Pareto-optimal scoring. This is batch processing for prompt engineering, not real-time adaptation.

Three-tier memory separates persistent facts (bounded markdown files like MEMORY.md and USER.md), searchable session history (SQLite with FTS5 full-text search), and pluggable external providers. The manual documents how these layers interact: the markdown tier loads into the system prompt on every session start, the FTS5 tier handles cross-session recall, and external providers handle specialized storage.

Self-evolving skills are the most unusual feature. Hermes creates skills from experience, refines them through a skill_manage tool, maintains them with a background Curator process that archives underperforming skills, and optimizes them offline through GEPA. Skills are not static packages installed from a marketplace. They mutate.

The OpenClaw Comparison

The manual includes a direct architectural comparison table between Hermes and OpenClaw. The framing: “Hermes packages a gateway around a learning agent; OpenClaw packages an agent around a messaging gateway.”

The differences are structural. OpenClaw supports 50+ messaging channels to Hermes’s 20+. OpenClaw’s skill ecosystem is larger but static. Hermes’s skills self-evolve. OpenClaw uses plain markdown files for memory. Hermes uses three tiers with full-text search.

On security, the manual notes that OpenClaw’s “gateway-first design and a large unvetted skill pool have been associated with publicized incidents in 2026,” while Hermes uses snapshot-before-write for file operations and a curated skill set. The manual adds a caveat: treat this as directional, not a current audit.

The manual also documents a direct migration path from OpenClaw to Hermes, suggesting Nous Research views OpenClaw’s user base as a conversion opportunity.

Why This Manual Matters for the Agent Ecosystem

Hermes has been shipping frequent updates since February, but documentation lagged behind release cadence. The manual corrects several inaccuracies from earlier informal guides: there is no “Hermes Vault” feature (credential management uses hermes auth), there is no /skill command (skills load via /<skill-name>), and built-in memory is file-based, not database-backed at the first tier.

According to a hands-on comparison published the same day on Medium, developers running both frameworks on constrained hardware (Raspberry Pi 4) found Hermes “felt more stable than OpenClaw” and “less confusing” after initial setup, though the author acknowledged prior experience with OpenClaw may have influenced that assessment.

The Self-Improvement Question

The deeper signal is architectural. OpenClaw’s 145,000+ GitHub stars represent the largest open-source agent community, but its skills are human-authored and static. Hermes is betting that agents should write, test, and optimize their own capabilities over time, with human operators supervising rather than authoring. The GEPA pipeline and background Curator process are implementations of that bet.

Whether self-evolving skills produce better outcomes than a large human-curated skill marketplace is an open question. The manual doesn’t claim to answer it. What it does is make the entire Hermes architecture legible enough for enterprise teams to evaluate the tradeoff on their own terms.

The Hermes Agent framework is MIT-licensed and runs on Linux, macOS, WSL2, Windows (early beta), and Android via Termux.