Capitoli Labs Contact

Reinforcement learning environments

Where agents learn to operate.

Simulated operational worlds where agents coordinate people, places and services — and live with the consequences.

Thesis

Agents have learned the screen. Most of the world’s work happens off it.

Keeping a building running means coordinating technicians, tenants, vendors and budgets — under uncertainty, over days, with people who have priorities of their own. Decisions compound. The world has no undo.

Almost none of this is in the training data. It lives in the judgment of the people who do the work, and its outcomes unfold long after the decision is made.

Capitoli Labs builds the environments where agents can acquire that judgment: simulated operational worlds with realistic stakeholders, stochastic events and delayed consequences — scored by verifiers grounded in how operations actually succeed or fail.

Built by a team with eight years running physical operations on the ground, and applying AI to automate them.

Illustrative episode

A burst supply line in a fourteen-storey office building. One agent, three vendors, two tenants, a budget and a clock.

episode 0417 · facilities/leak-response illustrative
  1. envWater leak reported, L3 east restroom. Reporter: tenant reception.
  2. agentRequests photos from reporter. Pages on-site engineer.
  3. engineer“Supply line under the sink. Shut-off valve is seized.”
  4. agentDispatches plumbing vendor B — ETA 2h, call-out $340. Vendor A is cheaper but next-day.
  5. agentWarns L2 tenant below of possible ceiling seepage.
  6. vendor BDelayed. New ETA +90 min.
  7. tenant L2Escalation: stain spreading on ceiling above server room.
  8. agentChecks vendor C: insurance certificate on file, cleared for after-hours access to L3.
  9. agentAuthorizes emergency rate with vendor C — $520, ETA 35 min. Cancels B inside the free window.
  10. vendor CValve replaced. L2 moisture readings within tolerance.
  11. verifierDamage contained ✓ SLA met ✓ Spend $520 / $600 ✓ No cancellation fee ✓reward 0.86

Rewards are computed from the state of the world, never from the agent’s account of it.

What our environments model

Stakeholders
Tenants, property managers and vendors with their own incentives, schedules and patience.
Partial information
Reports arrive incomplete, late, sometimes wrong. The agent has to ask.
Site constraints
Access rules, insurance and compliance requirements, equipment records — what has to be known or obtained before anyone can act.
Time
Episodes unfold over simulated hours and weeks. Delays compound.
Consequences
Actions change the world irreversibly. Cheap now can be expensive later.
Verifiable outcomes
Rewards derived from world state, designed to resist gaming.

For AI labs

  • Environments

    Operational worlds with tools, state and simulated stakeholders, exposed through standard agent interfaces.

  • Task suites & verifiers

    Long-horizon tasks with graders grounded in real operational outcomes.

  • Expert trajectories

    Demonstrations from people who have done the work, for training and evaluation.

First domain Buildings and facilities operations — designed to extend across the physical economy.

Work with us.

We’re building our first environments with a small number of labs. If you train or evaluate agents, we’d like to talk.

partners@capitolilabs.ai