Today's agents are good for an afternoon. Over weeks they lose context, lose track of each other, and cost too much to keep alive. We work on every layer that breaks.
Simulation environments, agent swarms and inference, built toward one goal: agents that keep working, and keep their memory, for a month.
Work with the LabA month-long agent fails in three places. It never trained on problems where actions play out over days. It can't coordinate with the agents beside it. And holding all that context in memory gets too expensive to run. Each track takes one of those on.
01 · Where agents learn
Digital twins of real operations, rebuilt from public records: rail networks recovering from real disruptions first, then power-grid redispatch. The agent replans as the day unfolds, and every plan is graded against a proven optimum.
02 · How agents run
Swarms of up to 10,000 agents that share state through a REPL, as in Recursive Language Models. Agents hand off work as variables, and every result can be checked by code.
03 · What it costs
Every agent's memory lives in the KV cache. We study how KDA, CSA, FlashAttention, vLLM and RadixAttention shape it, and what it takes to keep thousands of agents running at once.
How we build an environment
Most RL environments for language models are puzzles, code or web tasks. Real operations look different: someone keeps a schedule running under strict rules while things keep going wrong. We rebuild that work from the records it leaves behind, and grade agents against answers that can be proven best.
Public timetables, incident logs and operator records give us the real network, the real traffic and the plan that was actually run.
Breakdowns, closures, delays and weather, modelled on incidents that really happened, stacked by difficulty from routine days to crises.
An exact solver computes the optimal plan for each task. Tasks a simple heuristic already solves are dropped, so every task left is hard.
We replay what human dispatchers actually did through the simulation. If the twin can't reproduce the real outcome, it isn't ready to train on.
News reaches the agent over time, not all at once. Decisions lock in as the clock moves, so the agent has to plan, commit and replan for hours or days.
Rules first, then cost against the proven optimum. No LLM judge, no sampling: the same plan always gets the same reward, which is what RL needs.
What is real and what is simulated
We say it plainly for every environment. Timetables and incidents come from records. Where the records stop, such as exact capacities or confidential network data, we model it and label it as modelled. Rail comes first because its operating data can be checked end to end. Power grids follow, with their limits stated.
A real system, rebuilt from its records, that unfolds over days and grades every decision.
Thousands of agents work the problem, sharing state through code instead of context windows.
A KV cache small and fast enough to keep every one of them running.
The thread through all three is memory: in the world an agent acts on, across the agents in a swarm, and on the GPU that runs them.
Before the current tracks, we built verified environments and evaluations for cyber agents. The same discipline carries over.
All research notes
