Guide · Agent skill
How to use the Hermes Ephemeral Delegation agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Hermes Ephemeral Delegation is an agent skill that teaches an AI coding agent when to hand part of a task to a temporary helper agent instead of doing everything in one conversation. This guide is for developers who have not used agent skills before. By the end you will know what the skill changes, how to install it, how to phrase requests so it triggers, and what to check before trusting its results.
Hermes Ephemeral DelegationGentleman-Programming/gentle-ai · MIT Download the skill Quick start
- 1Download the ZIP. Download the skill ZIP from this page and unzip it so the folder contains SKILL.md.
- 2Place the folder. Move the unzipped folder into ~/.claude/skills/ for personal use, or into the project's .claude/skills/ folder for one repository.
- 3Or upload in the Claude apps. If you use the Claude apps, upload the skill under Settings → Capabilities → Skills.
- 4Confirm delegate_task exists. Make sure your agent runtime provides a delegate_task tool, since the skill cannot delegate without it.
- 5Ask with a named reason. Describe the task and why it should be split, such as mapping a large codebase or reviewing a diff with fresh eyes.
- 6Verify the output. Check file changes, test results and any URLs or IDs the worker reports before accepting the result.
What is an agent skill and what does this one make the agent do?
An agent skill is a folder containing a SKILL.md file of instructions that an AI agent loads when a task matches the skill's description. This one is written for the agent acting as the parent orchestrator in a conversation.
The skill tells the agent to delegate only when there is a named reason, such as mapping a large codebase, running two or more independent pieces of work together, reviewing a risky diff with fresh eyes, or debugging across many steps. For each delegation the agent writes a self-contained mission and starts a worker with a clean context. Small or simple work stays inline. The agent is also told to check worker claims, such as file writes and test results, before reporting success.
- Delegation is for a reason, not for size: a large task alone is tracked, not delegated.
- Workers are ephemeral and start with no memory of the parent conversation.
- Worker output is treated as a self-report and verified before it is passed on.
How do you install the Hermes Ephemeral Delegation skill?
Download the ZIP for this skill from this page and unzip it. The folder must contain SKILL.md. There are two common places to put it for Claude-based tools: a personal skills folder at ~/.claude/skills/, or a project folder at .claude/skills/ inside your repository, which keeps the skill with that project.
If you use the Claude apps instead of a folder on disk, upload the skill under Settings → Capabilities → Skills.
- Personal use across projects: place the unzipped skill folder in ~/.claude/skills/.
- Single project only: place it in the project's .claude/skills/ folder.
- Claude apps: upload it under Settings → Capabilities → Skills.
Check that your runtime supports delegate_task
The skill depends on a delegate_task tool that launches worker agents. If your agent runtime does not provide this tool, the skill cannot perform delegation, so check this before you rely on it. The source does not list which specific products provide delegate_task, so confirm it in your own tool setup.
Workers can only use the toolsets, MCP servers or skills named in their mission, as configured by you. The skill does not give workers access to anything extra.
How do you trigger the skill with natural requests?
You do not need special syntax. Describe the work and name the reason for splitting it, so the agent can match one of the skill's triggers: mapping, parallel units, context backstop, high-risk verification, fresh review, or multi-step debugging. Vague requests may leave the agent working inline, which is the intended behaviour for small tasks.
The examples below are illustrations of phrasing, not commands the skill requires.
- Mapping: Map where authentication logic lives across this large repo before we decide where the change goes.
- Parallel units: Update the parser code, its tests and the docs as three independent pieces and verify each one.
- Fresh review: Review this diff with fresh eyes before I open the pull request.
- High-risk change: This migration is risky, so have a separate worker verify it after you make the change.
- Multi-step debug: Work through this flaky test debugging in steps and keep my main conversation short.
What settings control workers?
The skill ships a tuning reference, references/tuning-knobs.md, that lists the delegation settings. These are set in ~/.hermes/config.yaml under the delegation key, or passed per delegate_task call. Defaults in that file include a maximum delegation depth of 2, at most 4 workers active at once, and no automatic approval of tool calls.
Setting max_spawn_depth to 1 prevents workers from starting their own workers. Leave subagent_auto_approve at its default of false unless you have a specific reason to change it, because the default makes workers prompt for approval on dangerous calls.
- max_spawn_depth (default 2): set to 1 to stop nested workers.
- max_concurrent_children (default 4): caps parallel workers.
- inherit_mcp_toolsets (default false): when false, list toolsets explicitly in the mission.
- child_timeout_seconds and max_iterations: per-worker limits, defaulting to the agent default.
What about safety: scripts and network?
The listing states that this skill runs no scripts and uses no network access itself. Its effects come from the worker agents it starts, which can use whatever toolsets, MCP servers or skills you have configured and named in the mission. Review those permissions before you let an agent delegate on a project with sensitive data.
Do not ask the agent to create persistent agent files or profiles. The skill says workers are ephemeral and must not be given persistent agent files.
- No scripts are bundled according to the listing.
- No network use is declared by the skill.
- Keep subagent_auto_approve off so dangerous worker calls prompt you.
- Check worker results yourself, especially file writes, test outcomes and URLs.
What people use it for
- Mapping a large codebase to decide where a change belongs without filling the main conversation with file reads.
- Running independent code, test and documentation updates in parallel and verifying each separately.
- Getting a fresh-context review of a diff before opening a pull request.
- Debugging a multi-step problem in a worker so the parent conversation stays small.
Frequently asked questions
Is the Hermes Ephemeral Delegation skill free?
The listing says the repository is under the MIT licence. The SKILL.md file's own header names Apache-2.0, so check the LICENSE file in the source repository before you rely on either one for commercial use.
What do I need to run the Hermes Ephemeral Delegation skill?
You need an agent runtime that provides a delegate_task tool for launching worker agents. Workers can also use any toolsets, MCP servers or skills you have configured and named in their mission.
Where do I install an agent skill like this one?
Put the unzipped skill folder in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder. In the Claude apps, upload it under Settings → Capabilities → Skills.
Why isn't my agent delegating the task?
The skill only delegates when a named reason applies, such as mapping, parallel units, a risky change, fresh review or multi-step debugging. Small or simple tasks are kept inline by design, and a large task alone does not count as a reason.
Does this skill run scripts or use the network?
The listing says it runs no scripts and uses no network itself. Any network or tool access comes from the toolsets and MCP servers you give to worker agents.
How does this compare to just asking the agent to do the task in one chat?
A single conversation keeps all context in one place, which is simpler for small work. This skill is useful when splitting work keeps the main context small or lets independent parts be checked separately, but workers start fresh and cannot share memory across steps.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you work on large codebases with an AI coding agent, need to keep the main conversation small, or want an independent fresh-context review of a risky change, and your runtime provides delegate_task. Pick something else for small or simple tasks, or for workflows where workers must share memory across steps, since the skill itself notes that delegation adds cost without benefit in those cases.
Download the skill