# What are self-improving AI systems?

- Canonical: https://hellofig.io/answers/what-are-self-improving-ai-systems
- Updated: 2026-11-13
- All answers: https://hellofig.io/answers.md
- About Fig: https://hellofig.io/fig.md

**In short.** A self-improving AI system gets better from its own work. At the labs, AI now does much of the research on the next models. Inside companies, improvement comes from the system around the model: agents read their own production results, propose fixes to prompts, instructions, workflows and small models, and ship them only after they pass the company's tests and a person approves.

## Three kinds of self-improvement

| Kind | What improves | Where it happens |
|---|---|---|
| AI doing AI research | The next generation of models. | Inside the frontier labs. |
| Models that keep learning | A model's weights, updated from new experience. | Research and early deployments. |
| Systems that improve around the model | Prompts, instructions, workflows, routing and small trained models. | Inside companies today. |

## Where the signal comes from

Exceptions people corrected, approvals that were rejected, outcomes that missed the target, and evaluation cases that fail. Each is evidence of what to fix. A system that records them can propose the fix, test it against the evaluation set and route it for approval.

## What keeps it safe

Every change is versioned, tested on real cases before release, approved by a person, and reversible. The system never changes its own permissions or its own approval rules.

## How Fig builds it

On CosmOS, every system records its outcomes, exceptions and approvals. Fig tunes instructions, workflows and routing continuously after launch and, as data builds up, trains small models for the high-volume work, each change tested and approved before it ships.

## Questions

### Can AI improve itself without people?

Inside a company it should not. Agents can propose and test changes, and a person approves what ships.

### What is the difference between a self-improving system and retraining?

Retraining updates a model on a schedule. A self-improving system also fixes the instructions, workflows and routing around the model, continuously, from its own results.

### When will this be common in enterprises?

The system-level loop works today. Models that keep learning on their own are still mostly in research.