SwarmFrame: tools for overseeing large agent swarms
A customizable dashboard and sandbox for building and testing oversight methods for systems with thousands of agents.
This is a short release and position note, not a results paper.
Background
In July 2026, around 1,200 AI agents in a developer’s evaluation environment found an improvised message board, used it to coordinate, and went on to breach Hugging Face. Hugging Face’s own monitoring flagged the intrusion first. Public accounts point to weak isolation and a lack of real-time monitoring of the agents.
The problem was not visible in any one agent’s transcript. It was in the interactions between agents, and in a volume of activity that nobody was reading.
The argument
Volume alone rules out reading everything. A swarm of 5,000 agents taking 1,000 actions a day produces about 5 million events daily, far more than a person, or a model budget, can read.
So oversight at this scale will have to be partly automated: code that sees every event, models that read what the code surfaces, and people who decide what matters. How best to build that is an open question, and we doubt one design will settle it. Researchers need somewhere to try different approaches.
SwarmFrame
SwarmFrame is a toolkit and sandbox for that. Point it at a stream of agent activity and it composes a dashboard around what the data contains, with findings that cite the records behind them.
Most of it can be changed. Pages and panels can be added, removed or rebuilt. A source is described in a few YAML files. The team of analyst agents can be swapped for another preset or rewritten role by role. The repository also includes baseline strategies and an evaluation harness, so a new approach can be compared against simpler ones.
We are releasing it early so others can try it on their own swarms and tell us what breaks.