# Unsupervised tool evolution boosts data‑lake analytics agents

> The TEAL system improves LLM analysts on KramaBench and LakeQA without any human labels.

Oossa · 2026-10-21 · https://oossa.com/en/unsupervised-tool-evolution-boosts-data-lake-analytics-agents

Researchers released TEAL, a method that lets coding agents automatically build reusable tools while answering data‑lake questions. Two helper agents suggest tools—one scans the raw data, the other watches the main agent’s steps. A tool is kept only if the main agent uses it without extra work and the answer stays consistent. The team also added TEAL‑Q to make fake questions when real ones are scarce.

## The facts

- Paper posted on 2026-10-21
- TEAL lifted Sonnet‑5 by 9.3 points on KramaBench and 21.7 points on LakeQA with a 15‑step limit

## Why it matters

It shows that analytics agents can become more efficient without any human‑provided labels, lowering the cost of building data‑lake tools.

## Sources & references

1. [Unsupervised Tool Evolution for Coding Agents in Data-Lake Analytics](https://arxiv.org/abs/2610.27327) – arXiv, 2026-10-21

Last updated: 2026-10-21