/harness-subagent
Use this skill to create and manage subagent entities in the agent-harness workspace. A subagent is a specialized agent definition — with a name, description, optional tool restrictions, and an optional model override — that harness renders into provider-native agent configuration files for every enabled provider.
What is a subagent entity?
A subagent entity is a single canonical Markdown source file under .harness/src/subagents/<subagent-id>.md. It contains:
- Required YAML frontmatter:
name and description
- Optional frontmatter: provider-specific fields set via override sidecars (model, tools, handoffs)
- Body: Markdown prose that becomes the agent's system prompt
Harness renders one output artifact per enabled provider from that single source file. You edit the source once; all provider files stay in sync.
Provider output mapping
Claude Code
| Item | Value |
|---|
| Output path | .claude/agents/<subagent-id>.md |
| Format | Markdown with YAML frontmatter |
| Required frontmatter | name, description |
| Optional frontmatter | tools (string or string[]), model |
Claude Code loads agent files from .claude/agents/ at session start. Each file's description field tells Claude when to delegate to that agent. Claude routes tasks automatically, or the user can invoke an agent explicitly with @agent-<name> or the /agents command. Supported model values: sonnet, opus, haiku, a full model ID such as claude-sonnet-4-6, or inherit (default). Subagents run in isolated context windows and cannot spawn further subagents.
Official docs: https://code.claude.com/docs/en/sub-agents
OpenAI Codex CLI
| Item | Value |
|---|
| Output path | .codex/config.toml |
| Format | TOML, section [agents.<subagent-id>] |
| Required fields | description, developer_instructions (body) |
| Optional fields | model, model_reasoning_effort, nickname_candidates |
Harness writes all enabled subagents into .codex/config.toml under [agents.<id>] entries. The description drives when Codex routes to that agent role; the body becomes developer_instructions. Codex also supports standalone TOML agent files in .codex/agents/ (outside of harness management), which are auto-discovered and take precedence over same-named inline entries.
Official docs: https://developers.openai.com/codex/config-reference
GitHub Copilot
| Item | Value |
|---|
| Output path | .github/agents/<subagent-id>.agent.md |
| Format | Markdown with YAML frontmatter |
| Required frontmatter | description |
| Optional frontmatter | name, tools (string[]), model, handoffs, mcp-servers, target, disable-model-invocation, user-invocable |
Copilot discovers agent files in .github/agents/ at the repository level. The handoffs field is Copilot-specific and is supported in VS Code agent mode: it defines routing buttons that let users switch to another agent mid-conversation, optionally with a pre-filled prompt. Note: handoffs are currently not supported by the Copilot coding agent on GitHub.com — they are silently ignored there. Example handoff entry (VS Code only):
yaml
1handoffs:
2 - label: "Start Implementation"
3 agent: implementation
4 prompt: "Now implement the plan outlined above."
5 send: true
When send: true is set, the handoff prompt submits automatically.
Official docs: https://docs.github.com/en/copilot/reference/custom-agents-configuration
Provider-specific frontmatter comparison
| Field | Claude Code | Codex CLI | GitHub Copilot |
|---|
name | required | via TOML key | optional |
description | required | required | required |
tools | string or string[] | — | string[] |
model | optional | optional | optional |
developer_instructions | — | required (body) | — |
handoffs | — | — | optional (VS Code only) |
Provider-specific options (model, tools, handoffs) are set through override sidecar files, not in the canonical source frontmatter — see the Override sidecars section below.
markdown
1---
2name: <human-readable display name>
3description: <when this agent should be used — used by all providers>
4---
5
6You are a [role description]. When invoked:
7
81. [Step one]
92. [Step two]
103. [Step three]
11
12Focus on [specific domain]. Do not [out-of-scope actions].
Only name and description are required in the canonical frontmatter. The body becomes the agent's system prompt for all providers.
Override sidecars
To set provider-specific options (model, tools, handoffs), create or edit the generated override YAML files:
.harness/src/subagents/<subagent-id>.overrides.claude.yaml.harness/src/subagents/<subagent-id>.overrides.codex.yaml.harness/src/subagents/<subagent-id>.overrides.copilot.yaml
Example Claude override sidecar:
yaml
1version: 1
2options:
3 model: haiku
4 tools:
5 - Read
6 - Grep
7 - Glob
Example Copilot override sidecar:
yaml
1version: 1
2options:
3 model: gpt-4o
4 tools:
5 - search
6 - fetch
7 handoffs:
8 - label: "Hand off to planner"
9 agent: planner
10 prompt: "Now create a plan based on the research above."
Harness CLI commands
bash
1# Scaffold a new subagent source file and register it in manifest.json
2npx harness add subagent <subagent-id>
3
4# Preview what files will be generated (dry run)
5npx harness plan
6
7# Write all provider artifacts from current canonical sources
8npx harness apply
9
10# Watch sources and auto-apply on changes
11npx harness watch
12
13# Remove a subagent entity and its source files
14npx harness remove subagent <subagent-id>
15
16# Remove entity but keep source files on disk
17npx harness remove subagent <subagent-id> --no-delete-source
Complete example
Source file: .harness/src/subagents/code-reviewer.md
markdown
1---
2name: code-reviewer
3description: Expert code review specialist. Reviews code for quality, security, and maintainability. Use proactively after writing or modifying code.
4---
5
6You are a senior code reviewer ensuring high standards of quality and security.
7
8When invoked:
91. Run git diff to see recent changes
102. Focus on modified files
113. Begin review immediately without asking for clarification
12
13Review checklist:
14- Code is clear and readable
15- Functions and variables are well-named
16- No duplicated logic
17- Proper error handling and input validation
18- No exposed secrets or API keys
19- Adequate test coverage
20- Performance considerations addressed
21
22Provide feedback organized by priority:
23- Critical issues (must fix before merging)
24- Warnings (should fix)
25- Suggestions (consider improving)
26
27Include specific examples of how to fix each issue. Do not modify files yourself.
Override sidecar for Claude (.harness/src/subagents/code-reviewer.overrides.claude.yaml):
yaml
1version: 1
2options:
3 model: sonnet
4 tools:
5 - Read
6 - Grep
7 - Glob
8 - Bash
After running npx harness apply, this produces:
.claude/agents/code-reviewer.md — used by Claude Code- An entry in
.codex/config.toml under [agents.code-reviewer] — used by Codex
.github/agents/code-reviewer.agent.md — used by GitHub Copilot
Typical workflow
bash
1# 1. Scaffold the subagent source
2npx harness add subagent code-reviewer
3
4# 2. Edit the source file
5# .harness/src/subagents/code-reviewer.md
6
7# 3. (Optional) Edit provider-specific overrides
8# .harness/src/subagents/code-reviewer.overrides.claude.yaml
9
10# 4. Preview the plan
11npx harness plan
12
13# 5. Generate provider artifacts
14npx harness apply
Check .claude/agents/, .codex/config.toml, and .github/agents/ to verify the rendered output.
References
