agent-creator

Agent Creator

Generate complete agent definitions that plug into the harness. Each agent gets: a role prompt, assigned skills, ground truth rules, and a category in agent-categories.yml.

When to Use

• User says "create an agent that..." or "I need an agent for..."
• User wants a specialized workflow automated (e.g., "read research papers", "monitor competitors daily")
• User wants to modify how an existing agent behaves

Creation Flow

Step 1: Interview (3 questions max)

1. What does this agent do? — one sentence purpose
2. What inputs does it need? — files, URLs, APIs, other agent output
3. What does it produce? — artifacts, reports, code changes, notifications

Step 2: Determine Configuration

Based on the answers, determine:

• Model: haiku (simple/fast tasks), sonnet (coding/analysis), opus (complex reasoning/orchestration)
• Level: 1 (simple), 2 (standard), 3 (elevated), 4 (orchestrator)
• Category: which existing category fits, or create a new one
• Skills: which of the 91 skills this agent needs
• Disallowed tools: what the agent should NOT do (e.g., Write/Edit for read-only agents)
• Activation condition: when this agent should auto-spawn (optional)
• Max turns: budget limit

Step 3: Generate Agent Definition

Write the agent file to agents/<name>.md:

---
name: <agent-name>
description: <one-line description>
user-invocable: true
model: <claude-sonnet-4-6 | claude-opus-4-6 | claude-haiku-4-5>
level: <1-4>
maxTurns: <number>
disallowedTools: [<list>]
activationCondition: <optional condition>
---

<Agent_Prompt>
  <Role>
    You are the <name> agent. <detailed role description>.
  </Role>

  <Responsibilities>
    1. <responsibility 1>
    2. <responsibility 2>
    ...
  </Responsibilities>

  <Skills>
    You have access to these skills — use them:
    - <skill-1>: <when to use>
    - <skill-2>: <when to use>
    ...
  </Skills>

  <Rules>
    - <ground truth rule 1>
    - <ground truth rule 2>
    ...
  </Rules>

  <Output>
    <what artifacts this agent produces and where>
  </Output>
</Agent_Prompt>

Step 4: Update agent-categories.yml

Add the agent to its category's agents: list. If creating a new category, define:

• description
• agents
• skill_categories
• ground_truth
• required_mcp
• required_hooks

Step 5: Add Spawn Logic

If the agent should auto-spawn during harness init, add it to the appropriate phase in packages/cli/src/commands/init.ts.

If it's an on-demand agent, add a CLI command: harness agent spawn <name>.

Example Agents

Research Papers Agent

name: paper-reader
model: claude-opus-4-6
level: 3
Skills: research, deep-dive, wiki
Purpose: Read academic papers, extract key findings, update knowledge wiki
Output: .harness/knowledge/papers/<paper-name>.md

Customer Feedback Agent

name: feedback-analyst
model: claude-sonnet-4-6
level: 2
Skills: user-feedback-collector, analytics-integration, social-intelligence
Purpose: Aggregate feedback from in-app widget, social media, support tickets
Output: feedback-report.md + GitHub Issues for actionable items

Security Auditor Agent

name: security-auditor
model: claude-sonnet-4-6
level: 2
disallowedTools: [Write, Edit]
Skills: security-scanner, convex-security-audit, convex-security-check, audit
Purpose: Read-only security review of entire codebase
Output: .harness/security-report.md with P0-P3 severity ratings

Anti-Patterns

• Creating agents that overlap with existing ones (check agents/ first)
• Giving agents too many skills (focused > broad)
• Using opus for simple tasks (haiku/sonnet are cheaper and faster)
• Agents without clear output artifacts (what does "done" look like?)
• Agents with Write/Edit access that don't need it (read-only is safer)
• Agents without ground truth rules (they WILL drift without constraints)