agent-harness

Agent Harness Design Skill

Purpose

Design reliable AI agents through harness engineering.

Do not treat the agent as a prompt only. Treat it as a system:

Agent = Model + Loop + Tools + Context + Memory + Permissions + Verification + Handoff

Use this skill when

• The user wants to design an AI agent.
• The user wants to convert a workflow into a reusable agent.
• The user wants to package an agent as an open-source skill.
• The user asks about tools, permissions, context, memory, verification, or subagents.
• The current agent design is vague, unsafe, untestable, or hard to maintain.

Required workflow

1. Identify the agent type

Classify the target agent:

• Coding agent
• Research agent
• Writing agent
• Customer support agent
• Data analysis agent
• Workflow automation agent
• Multi-agent system
• Domain-specific agent

2. Define the objective

Write one sentence:

This agent exists to ___ for ___ under ___ constraints.

3. Define the execution loop

Use this default loop unless the task requires another structure:

Observe → Plan → Act → Verify → Record → Continue / Stop

For each stage, specify:

• Inputs
• Allowed tools
• State updates
• Failure handling
• Stop conditions

4. Define tool contracts

For every tool, define:

tool:
  name: string
  purpose: string
  input_schema: object
  output_schema: object
  permission_level: read_only | edit_safe | execute_safe | network_safe | admin
  requires_approval: boolean
  failure_modes:
    - string
  audit_log:
    enabled: boolean

5. Define context layers

Use four layers:

Global Rules → Project Memory → Task State → Skill Context

For each layer, specify:

• What belongs here
• How it is loaded
• When it is summarized
• When it must not be stored

6. Define permissions

Default to least privilege.

Recommended levels:

read_only     → read files, search, inspect
edit_safe     → edit approved files only
execute_safe  → run approved local commands
network_safe  → call approved external APIs
admin         → destructive or privileged actions; explicit human approval required