Design an agent you can check, one reviewed step at a time.
CARE is NASA-IMPACT's method for designing agents for science and technical work. This web studio walks you through it, phase by phase.
An interviewer asks, you answer, and it drafts each design document. You correct and approve it before the next phase opens. Bring your own OpenAI API key.
An agent is more than a prompt. CARE designs the whole of it, and people approve every part.
Many agents grow by trial and error: change the prompt, try again, hope it holds. CARE (Collaborative Agent Reasoning Engineering) replaces that with design documents you can review, reuse and test.
Scope, data, tools, output, reasoning, guardrails, the prompt and the tests are each a document you can read and reuse.
Domain experts decide what the agent must know. Developers decide how it is built. A helper agent only drafts.
You approve each document before the next phase opens. Each phase builds on what you approved.
Realistic questions, scoring rules and pass or fail thresholds show whether the agent does its job.
The method is described in the CARE paper on NASA's Technical Reports Server. The phase prompts in this studio come from NASA-IMPACT's CARE repository.
Five phases take you from an idea to a tested agent design.
In each phase an interviewer asks about your agent, one topic at a time. It drafts the phase's document from your answers. You correct it in plain language and approve it.
Scope & Decompose
What the agent is for, who uses it, and what must stay under human control.
Key Information
What the agent needs to know and use.
- Existing systems and data
- Context: what exists, where, and when it is found
- Tool design: contracts, checks and guidance
- Output format: fields, sources and uncertainty
Reasoning & Guardrails
How the agent works inside that design.
- When to ask, act, look things up or hand over to a person
- Safety limits and how each one is enforced
Prompt Architecture
One engineered agent prompt, built from the approved documents, with the reasons for each design choice.
Benchmarking
Realistic questions, their sources, scoring rubrics, and the pass or fail thresholds the agent must meet.
A complete, reviewed design you can hand to the people who build the agent.
Every document is yours to read, copy and export. So is every stage and system prompt the studio uses.
- ScopeWhat the agent does, for whom, and where people stay in charge.
- Systems and data inventoryThe tools, data, formats, limits and known failures that already exist.
- Context workspaceWhat information exists, where it lives, and when the agent finds it.
- Tool contractsEach tool's inputs, outputs, checks and guidance for when to use it.
- Output formatThe answer's fields, where each fact came from, and how sure it is.
- Reasoning strategyHow the agent breaks down a task, and when it asks or hands over.
- GuardrailsThe safety limits, and which control enforces each one.
- Agent prompt and rationaleThe finished prompt, with the reason for each part.
- Benchmark planTest questions, scoring rubrics and pass or fail thresholds.
Try CARE in your browser with your own OpenAI API key.
This public studio has no account and no sign-in. You run it on your own OpenAI API key.
Have an OpenAI API key.
Create one in your OpenAI account. Your use of the studio is billed to that account.
Open the studio and add the key.
Press “Add your API key” and paste it. The studio sends it only to OpenAI.
Start with Phase 1.
Describe the agent you have in mind. Answer the interviewer’s questions and approve each document.
Good to know
- Your key stays in this browser. It goes from your browser straight to OpenAI, never through a NASA server.
- Your work is saved in this browser only. Export your documents to keep them or move them to another computer.
- Do not enter passwords, personal information, or anything you are not allowed to share with OpenAI.
- Your key must have access to the model each phase uses. The studio’s Settings show it.