Research

Advancing the science of multi-agent autonomy

Our lab studies how agents plan, coordinate, and stay safe at scale. We publish openly because better autonomy is a problem worth solving together.

Focus areas

Where we push the frontier

Hierarchical planning

Decomposing open-ended goals into verifiable task graphs that recover from failure.

Agent coordination

Protocols for negotiation, delegation, and shared memory across heterogeneous agents.

Safe autonomy

Alignment, guardrails, and interpretability so agents stay within intent at scale.

Publications

Recent papers & releases

ZB-26-04

Mesh: Emergent Role Specialization in Agent Fleets

We show that uncoordinated agents converge on stable, efficient role divisions when given a shared memory bus and a lightweight market for tasks.

R. Okafor, L. Maren, D. Vasquez · zebrabeta Research · 2026
ZB-26-01

Verifiable Plans: Self-Checking Task Graphs for Autonomous Agents

A planning framework where every node carries its own success predicate, cutting silent failures by 71% across 40 benchmark workflows.

A. Iqbal, P. Sørensen, M. Chen · zebrabeta Research · 2026
ZB-25-09

Guardrail: A Programmable Policy Layer for Tool-Using Agents

An intermediary that intercepts every tool call against declarative policy, enabling safe autonomy without sacrificing throughput.

N. Adeyemi, J. Park · zebrabeta Research · 2025
ZB-25-03

Replay: Deterministic Tracing for Non-Deterministic Agents

A logging format that makes any agent run fully reproducible, enabling true debugging and post-hoc audit of autonomous decisions.

L. Maren, R. Okafor · zebrabeta Research · 2025
Open source

What we've released to the community

mesh-bench

An open benchmark of 120 multi-agent tasks with verifiable success criteria.

View on GitHub

traceformat

The open spec for deterministic agent traces, adopted by 30+ tools.

Read the spec

policy-dsl

A tiny declarative language for expressing agent guardrails as code.

Get started

Work on hard problems with us

We're hiring research scientists and engineers who want to define how autonomous systems work.