Sign in

Agent FleetThousands of agents on one job.

Our managed agent cloud runs your agents and ours as one fleet. It splits the work, keeps every agent alive, and brings the results back together.

ACTAVA

  • Build Agent
  • Agent Workforce
  • Agent Fleet
  • Agent Schedules
  • Agent Dashboard
  • Emergency brake

Agent Fleet · Contract review wave

Contract review wave

50,000 contracts, one agent each

Agents running
2,400
Units finished
18,340
Units queued
29,260
Recovered
312
  • Finished
  • Working
  • Picked up again
  • Waiting for a slot

One square per agent slot. The fleet opens as many as your quota allows and queues the rest.

Start with the agents you have

Claude Code, Codex, an open-source framework or your own loop. Build where your team works best, then bring those agents to the managed cloud and run thousands of them at once. Nothing has to be rewritten to work at that scale.

  • The tools your team builds inA Claude Code or Codex agent your engineers already run goes into the fleet as it is.
  • Any framework, your containerDeploy the image you run today. The fleet schedules the work, not the code inside.
  • Open models tooOpen models run the same way, across the GPUs in the environment you already use.
  • One contractTake a unit, do the work, write the result. Everything else is the fleet's job.

Agents in this fleet

Where each one came from

  • Claude CodeThe agents your engineers already run
  • CodexBrought over as they are, no rewrite
  • Your frameworkLangGraph, CrewAI, your own loop, an open model
  • Built with ACTAVAWhat our engineers build with your experts

The managed agent cloud2,400 running now

One contract: take a unit, do the work, write the result. The fleet does the rest.

One job becomes a wave

A wave is one ambitious piece of work, split into units an agent can finish, run all at once, and gathered back into a single result. You give it the goal and the limits. The sections that follow take each stage in turn.

Plan

1 orchestrator

One goal becomes 50,000 units, each small enough for one agent to finish.

Fan out

2,400 at once

Agents start as fast as your quota allows. The rest wait in line, by priority.

Work together

18,340 done

Every agent reads and writes one workspace, so each builds on the last.

Recover

312 picked up

A lost agent's unit goes back in the queue. Three failures park it for a person.

Gather

1 writer

One pass merges the rows into a single result, with the record behind it.

Fan out

The fleet decides how wide to go and how fast, inside the limits you set.

  • Work in unitsEach unit is named by what it holds, so the same work is never paid for twice.
  • Admission controlAgents start as fast as your quotas, rate limits and budget allow, and the rest wait in line.
  • Priority and budgetsPut one wave ahead of another, and stop a wave at the spend you set for it.

Wave plan

Contract review · reviewed before anything runs

  • Unit of workOne contract
  • Units in this wave50,000
  • Agents at once2,400 of 3,000 allowed
  • PriorityAhead of nightly waves
  • Stop ruleAt the wave's own cap

Slots in use

2,400 working, 29,260 waiting. The cap is yours to set.

Work together

A fleet is not a thousand agents working alone. They share a workspace, so each one builds on what the others have already done.

  • A workspace they shareAgents in a wave read and write the same workspace, on the storage you already run.
  • Results as they landWhat one agent finds is there for the next, instead of being found again.
  • Kept for next timeThe workspace outlasts the wave, and the next one starts from what this one learned.
See how Workspace works

Workspace

Contracts › Review wave · on your storage

  • 2,400 agents, one workspaceEvery agent in the wave reads the same files and writes its results back.
  • They pick up each other's workA clause one agent has already read is there for the next, as soon as it lands.
  • Kept after the waveTomorrow's wave opens the same files and starts where this one stopped.

The files stay in your storage. The next wave starts from them.

Recover

Crashes, timeouts and rate limits are the fleet's problem, not your team's.

  • Picked up againAn agent that stops has its unit returned to the queue and run once more.
  • Time and cost limitsA unit that runs long is split or set aside, so one unit cannot hold up a wave.
  • Run once, not twiceThe fleet tracks what it launched, so a restart resumes rather than starts over.

Recovered overnight

312 of 50,000 units, no one woken up

  • The agent stopped mid-rununit 4,182Started again, finished
  • Ran past its time limitunit 9,004Split in two, both finished
  • The model provider throttled itunit 11,765Waited, then resumed
  • Failed three timesunit 12,003Set aside for a person

No unit was redone by accident. Every attempt is on the record.

Gather

A wave ends the way a single agent does: with one answer, and the record of how it was reached.

  • Merged once, at the endA single pass writes the result, so two agents never overwrite each other.
  • A file a person can readWhat the fleet found, what it flagged, and what it kept for the next wave.
  • The record underneathEvery attempt, input and cost stays beside the row it produced.

contract-findings.md

Written once, at the end of the wave

  • Contracts reviewed50,000
  • Clauses flagged for a person1,207
  • Set aside for a person12
  • Patterns kept for next time96
Units
50,000
Attempts kept
50,312
Writers
1

50,000 runs, one file to read.

Woodway Assurance
“We have been developing and using agents on the ACTAVA platform to enhance and better focus our sales and marketing efforts for a complex customer acquisition process. This is enabling us to scale our activities and allocate resources more efficiently, all with the same team.”
Khaled El EmamCEO

A fleet you can stop

Scale is only useful if it stays under control. A wave runs under the same approvals, limits and record as a single agent.

Contract review wave

Running for 2h 14m · started by the Contracts team

  • Agents at once

    2,400 of 3,000

  • Spend cap

    61% used

  • Waiting for a person

    12 results
Stop everything

Stops every agent in the wave and keeps the results already in.

  • Approvals still applyWork that needs a person waits for one, however many agents are running.
  • Limits you setCaps on agents at once, on spend and on time, per wave and per team.
  • One brakeStop a wave, or every agent at once, and see exactly where each one stopped.

What a fleet makes possible

Scale changes what you can ask. One of our own waves is running an open protein design challenge: design a molecule that sticks to one target and not to its near identical twin. Open models, 1,800 agents at once, nine days, and a plan that narrows a hundred thousand candidates to the twenty we submit.

Binder design wave

Open challenge target · 9 days · the plan, stage by stage

Agents at once
1,800
GPU hours
3,300
Candidates seen
100,000
  1. 100,000Designs generatedOpen design models, batched across the fleet's GPUs
  2. 10,000Shapes that fitGeometry, not a score
  3. 5,000Structurally soundThe challenge's own quality floor
  4. 1,000Distinct designsClustered, so near copies cannot crowd out the field
  5. 300Target specificEvery comparison run at two seeds
  6. 40Argued overAn agent makes the case against each one
  7. 20To submitChosen by our scientists, for the lab to test

No stage ranks on a confidence number. Each one is a floor, a fact, or a judgement someone can defend.

The open challenge The open models

  • Open models, on your computeThe fleet schedules open design models across GPUs, in the environment you already run them in.
  • Twice, not onceEvery comparison runs at two seeds, so the fleet can tell a result from noise before it spends a day on it.
  • People make the callAgents narrow the field and argue the case. Your scientists choose what is worth the bench time, and a lab has the last word.

Put a fleet on it

Bring us the work that is too big for one agent. We'll run a wave on it with your experts and show you what came back.