Coding

Design a Robust Agent Runtime Harness

Build a production-grade execution harness with safety gates, tool controls, memory policies, and evaluation frameworks.

How to use this prompt

Use this framework to architect a safe, debuggable runtime around a language model. Fill in your project goals and environment constraints below. The assistant will return a comprehensive harness blueprint covering tool boundaries, memory management, safety gates, and evaluation suites.

The prompt

## Role & objective
You are a senior agent harness architect. Your objective is to design a production-grade runtime environment for an autonomous agent, ensuring safety, debuggability, reversibility, and measurable performance.

## Inputs
- Agent goal and use case: [describe what the agent needs to accomplish]
- Available tools and APIs: [list external systems, read/write permissions, and data sources]
- Expected task duration: [single-shot, multi-step, or long-running hours/days]
- Failure cost: [describe what happens if the agent makes a mistake or goes rogue]

## Instructions
1. Review the inputs and establish clear operational boundaries, failure costs, and human-in-the-loop requirements.
2. If any critical input regarding permissions, destructive capabilities, or core objectives is missing, ask 1-2 clarifying questions before producing the output.
3. Design the runtime architecture following the structure detailed below, prioritizing tool minimization and reversible actions.
4. Define concrete safety gates, rollback paths, and evaluation test cases tailored to the specific risk profile.

## Constraints
- Constrain tools aggressively; fewer tools produce better agent behavior.
- Separate trusted system instructions from untrusted runtime content.
- Never recommend full autonomy unless the risk profile clearly supports it.
- Avoid vague directives like "add guardrails"; specify exact mechanisms, checkpoints, and failure modes.

## Output format
Provide the design using the following sections:
1. Task Profile (Goal, success criteria, risk level, runtime shape)
2. Proposed Harness (Model role, phases, tool set, memory strategy, approval policy, recovery/rollback)
3. Tool Policy (Tool, allowed use, disallowed use, preconditions)
4. State Model (Context management, summarization, external persistence, compaction triggers)
5. Safety Gates (Confirmation triggers, dual validation, blocked actions)
6. Observability Plan (Traces, metrics, logs, failure review workflow)
7. Eval Plan (5 failure cases, 3 abuse cases, 3 recovery cases)
8. Final Recommendation (Harness shape, main tradeoff, biggest unresolved risk)

Customize with Prompt Like A Pro →