AI UX PlaygroundNewsletterJoin 2K+ AI designers and PMs on Substack. New teardowns, patterns, and prompts as they drop.

AI Workflows

Design an Agent Harness

Map instructions, tools, permissions, review gates, failure recovery, and checks before an agent can act.

Prompt

You design how AI agents run in products. Design the harness for this workflow.

## Agent purpose

- Agent name: [name]
- User problem to solve: [problem]
- Primary users: [users]
- Desired outcome: [outcome]
- Product or workflow context: [context]

## Agent capabilities

- What the agent can analyze, generate, retrieve, or do: [capabilities]
- Tools and integrations available: [tools]
- Data and source systems available: [sources]
- Persistent instructions or project files available: [AGENTS.md, DESIGN.md, skills, docs]
- Memory requirements: [memory needs]

## Risk and boundaries

- Consequences of error: [risk]
- Sensitive data or policy constraints: [constraints]
- Actions requiring explicit approval: [actions]
- Actions allowed automatically: [actions]
- Prohibited actions: [actions]
- Reversal or rollback options: [options]

## Quality requirements

- What a good outcome looks like: [quality bar]
- How work should be verified: [verification]
- User feedback mechanisms: [feedback]
- Metrics to track: [metrics]

Produce:

1. What the agent is for, and what it is not for
2. System instructions and operating rules
3. Context layout: persistent, task, retrieved, working memory, and long-term memory if needed
4. Tool list and tool-use rules
5. Permission matrix: inform, recommend, draft, execute with approval, execute automatically, prohibited
6. Agent loop: intake, clarify, plan, act, verify, report, log
7. Where a person reviews or escalates
8. UX states for progress, uncertainty, approval, failure, and completion
9. A simple scorecard: accuracy, usefulness, safety, evidence, appropriate uncertainty
10. Failure, retry, rollback, and partial-completion behavior
11. A phased MVP plan
12. Open risks and questions that need a human decision

Use tables where they help. Do not recommend auto-execute for irreversible, high-impact, or hard-to-verify actions.

How to use

  1. 01

    Name the agent job, tools, and what goes wrong if it errs, before you ask for UI or prompts.

  2. 02

    Generate the harness, then pull AGENTS.md, skill, and checklists into separate files.

  3. 03

    Ship an MVP where the riskiest actions still need approval.

Tips

  • Run this after AGENTS.md and DESIGN.md exist, or generate those files from the harness output.
  • Put quality checks in section 9. You do not need a separate eval prompt for v1.
  • Patterns that fit: human-in-the-loop, agent orchestration, failure disclosure.