Research
Lynwe studies what a group of agents needs in order to work like a team rather than a crowd: how they share what they know, how they divide work when they know different things, how they improve, and how anyone would tell whether a team is any good.
The hard problems in multi-agent systems are not inside the model. They are in what surrounds it.
What we are working on.
Some of it has results behind it. Some is still a question.
Shared memory
Five agents given one document read it five times and pay five times. Scoping memory by session, by agent, and by team brings that well below linear in our own runs.
Uneven information
Most systems fix who talks to whom before work starts, but the members of a real team do not all know the same things. Can the shape of a team follow the task instead?
Harness evolution
The scaffolding around a model is built by hand and then stops changing. Could an agent work on it itself? We have no answer yet.
Execution feedback
A sandbox returns whatever the process printed — a thousand log lines, a stack trace, an exit code. We are working on what turns that into something an agent can act on.
Beyond text
Almost all agent work so far has been text and code. Doing things needs somewhere to do them, so we are building environments an agent can act inside.
Team evaluation
Benchmarks were built for one agent over short runs. None of them says how a team holds up over a long one, or what it cost.