Researchers Introduce GAMPO Framework for AI Agent Governance

In a recent preprint published on arXiv, researchers introduced the Governed Autotelic Multi-Agent Product Organization (GAMPO) framework. Synthesized from 321 sources, the specification combines agency, agile practices, platform dynamics, and governance theory to structure autonomous agent operations.

The framework was tested on CHI-Bench, a long-horizon healthcare benchmark, across both open-source and frontier models. The results revealed that governance benefits are gated by a model's spare capacity. For capacity-constrained models, full governance structures yielded little improvement, whereas a simple verification instruction doubled task success.

On frontier models, replacing generic procedures with case-grounded definitions-of-done significantly boosted performance, lifting prior-authorization success up to 84% under self-consistency evaluation. The authors noted that these findings remain exploratory due to small sample sizes in the initial tests.