Ben Goertzel: Incrementally Building AGI with Agent Swarms

bengoertzel · x · 2026-08-08

Following the AGI-26 conference, SingularityNET founder Ben Goertzel shared his systematic thoughts on using agent swarms to build AGI. He observed many attempting to have agent swarms download AGI research papers and bash them into a working codebase, seeing the seeds of a workable AGI approach in this.

He detailed the core design philosophy of his OmegaHive project:

Guided Improvement Loop: Instead of blindly combining dozens of cognitive components, the system establishes a baseline and adds ONE cognitive mechanism at a time (e.g., uncertain reasoning, attention economics, predictive coding), which is then evaluated, tuned, and either promoted or parked.
Evaluation Ecology: Noting that traditional AGI scores get hacked, he built a complex test suite of 10 environments (e.g., AGI Maze, a 2D world harder than ARC-AGI-3; Neoterics, a 3D synthetic ecology; LeanGarden for theorem proving). A "Transfer Ring" measures knowledge flow across these environments.
Test Tuning: By training narrow AIs (ML, Bayesian, evolutionary) as baselines and test-tuners, the system ensures tests are neither too easy nor too hard, and lack shortcuts.
Human-in-the-loop: Instead of typing code, human researchers answer questions at stuck-points, offering the most efficient use of researchers.

He envisions 10,000 self-uplifting OmegaHives eventually connecting via OmegaBuzz into a decentralized global brain.

Original post →

Ben Goertzel: Incrementally Building AGI with Agent Swarms

More from AGI Musings