Breaking

UCSB Releases Self-Evolving Agent Framework GEA, Hits 71% on SWE-bench

July 10, 2026 at 13:32 EDT

AI Agents
Research & Papers
Software Dev & Coding

A research team led by the University of California, Santa Barbara (UCSB) has released the implementation of GEA (Group-Evolving Agents), a framework in which multiple AI agents autonomously keep improving by sharing experience, on GitHub. It scored 71.0% on the software-engineering benchmark SWE-bench Verified, reaching a level comparable to human-engineered systems through self-improvement at no added cost.

February 2026 · GEA — Group-Evolving Agents

Self-Improving AI Agents Reach Human-Engineered Level — at Zero Added Cost

GEA treats the whole group of agents as one evolving unit, sharing experience so the collective keeps improving. It hits 71.0% on SWE-bench Verified — matching top human-built systems through self-improvement alone.

71.0%

SWE-bench Verified — matches human-engineered SOTA (71.8%)

88.3%

Polyglot multilingual generation — new best

1.4

avg. iterations to fix a framework bug (vs 5 prior)

SWE-bench Verified — GEA vs the prior self-improving best

Higher column = higher score. Both measured on the same benchmark.

56.7%

Prior self-
improving SOTA

71.0%

GEA
(self-improving)

71.8%

OpenHands
+ GPT-5 (human)

How the group evolves — experience sharing loop

Agents explore & solve tasks

→

Share experience across the group

→

Select next parent by Performance–Novelty

→

Group improves cumulatively

Why it matters — and what's still unknown

The group — not the individual agent — is the unit of evolution, turning exploration diversity into cumulative gains.
Improves with less human intervention, hinting at reduced reliance on prompt engineers.
Only research code is released — no commercial version or pricing confirmed, and real-world evaluation is still to come.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.