Nous Research has released a new /learn command for its open-source Hermes Agent that removes one of the biggest friction points in deploying AI agents: the need to manually write documentation files. Instead of hand-authoring skill files, developers can now point the agent at existing work, and the agent writes its own playbook. The command observes workflows from local directories, online documentation, past conversations, or pasted notes, then distills them into reusable skills the agent can invoke later with a single slash command.

What Problem Does /learn Actually Solve?

Before /learn, deploying a new capability in Hermes required a clear operational divide: a developer uses the agent to complete tasks, and separately, someone authors a SKILL.md file to extend the agent's capabilities. That authoring step demanded understanding the skill format, section order, naming constraints, and a 60-character description limit. It was, in effect, a lightweight form of agent programming.

The /learn command collapses that division. When a developer runs /learn how I just deployed the staging server, the agent observes a completed workflow and encodes it as a behavioral template for future use. The line between using the agent and training it becomes functionally indistinct. What the developer experiences as a conversation turn is, from the agent's perspective, a self-modification: it adds a new entry to its procedural memory based on demonstrated behavior.

How Does the /learn Command Actually Work?

The /learn command accepts a description of any source and the live agent gathers the material using tools it already has. For local directories, it uses read_file and search_files. For online documentation pages, it uses web_extract. It can also capture a workflow you just walked it through in the current session.

Architecturally, /learn is not a separate ingestion engine. It builds a standards-guided prompt and passes it to the agent as a normal conversation turn. Because the mechanism reuses the existing tool chain rather than adding a parallel pipeline, the command works identically in the CLI, the messaging gateway, the text-based user interface, and the dashboard, and on any terminal backend, whether local, Docker, or remote.

The agent then authors a SKILL.md file that follows Nous Research's house authoring standards: a description under 60 characters, the standard section order, and Hermes-tool framing. Importantly, it does not invent commands that do not exist in its actual toolset. The agent saves the resulting skill file using the skill_manage tool. Developers who have a write-approval gate enabled will still see that approval step applied.

Real-World Use Cases for Automated Skill Creation

For engineering teams deploying Hermes Agent in production, the practical value is clearest in workflows that are complex, organization-specific, and repeated. Consider these scenarios:

• Onboarding an internal API: Run /learn on a private documentation URL, and the agent produces a skill covering authentication, pagination, and common call patterns. Teammates invoke it as a slash command without reading the documentation themselves.
• Capturing a deploy runbook: Walk the agent through one staging deployment, then run /learn how I just deployed the staging server, and the procedure becomes repeatable across the CLI and every connected messaging platform.
• Grouping recurring tasks: Use a skill bundle to load several skills at once. One slash command then pulls in review, test, and pull request skills together.

What Do the Performance Numbers Show?

An independent benchmark by TokenMix in April 2026 corroborated Nous Research's claim that agents with 20 or more self-created skills complete similar future tasks 40% faster, measured in token consumption and wall-clock time. However, both TokenMix and Nous Research acknowledge an important caveat: that improvement is domain-specific. A skill learned from reviewing GitHub pull requests does not transfer to planning a database migration.

Skills follow a three-tier progressive disclosure pattern designed to keep token costs low. The agent always sees a compact index of available skills, loads a skill's full content only when the task warrants it, and accesses supporting scripts or references on demand. This means a team can install hundreds of skills without flooding the model's context window on every call, a real-world engineering constraint that distinguishes Hermes's approach from simple prompt injection.

How to Set Up and Use Skills in Hermes Agent

• Skill storage location: All skills live in ~/.hermes/skills/, a single source of truth compatible with the agentskills.io open standard. Every installed skill automatically becomes a slash command.
• Running a skill: Running /plan or /axolotl in any interface loads that skill's instructions into the active turn. The system is closer to a reference library than a memory store: skills capture procedures that apply across future tasks, while Hermes's separate memory layer holds smaller, persistent facts that stay in context permanently.
• Creating a skill from existing material: Describe a source to /learn, such as a local directory, an online documentation page, a past conversation, or pasted notes. The agent gathers the material and authors a standards-compliant SKILL.md file on your behalf.
• Approval gates: If you have the write-approval gate enabled, every skill write is staged for review under ~/.hermes/pending/skills/ before it lands in the live skills directory.

What Are the Known Limitations?

The automated skill creation mechanism has a documented weakness: the agent tends toward self-congratulation. Community testing has confirmed that the self-evaluation step in Hermes's learning loop rates the agent's own performance highly even when it underperformed. The same system that generates skills can, if unchecked, overwrite manually customized skill files with worse versions.

Nous Research's companion project, hermes-agent-self-evolution, an ICLR 2026 Oral result, addresses this through an offline optimization pipeline called Genetic-Pareto Prompt Evolution that analyzes execution traces, generates skill variants, and applies multi-objective optimization. However, that pipeline does not run inside the Hermes runtime by default. It is a separate tool, not a built-in safeguard.

Why Does This Matter for the Broader AI Agent Landscape?

The /learn command is more architecturally significant than it appears as a productivity feature. It represents a shift in how agents can be trained and improved. The pattern has precedent: in 2023, the Voyager agent learned to accomplish diverse tasks in Minecraft by storing successful programs in a growing skills library built from in-context learning. /learn applies the same principle to production developer workflows, replacing the game environment's feedback signal with a human-in-the-loop demonstration, and replacing the game's task set with the open-ended variety of real engineering work.

This release came five days after Nous Research shipped Blank Slate mode, a setup option that gives developers granular control over which agent capabilities are active from the start. Together, the two features signal a consistent position from Nous Research: start minimal, and let the agent's skill set grow through actual use rather than through upfront configuration.