const harness = createHarness({ tenant }) const lessons = await harness.recall(task) for (const step of run.steps) { await capture(step, { experience: run.id }) } await harness.refine({ outcome: "merged" }) export const agent = defineAgent({ model: "route:builder", tools: [repo, shell, browse], learning: { lessons: 6, capacity: 30 }, }) allocate deploy --agent builder packing instruction.md + skills/ digest 9f2c41e · head moved signal bound e_4821 ground_truth signal bound e_4822 execution compaction 1,023 experiences → adapter

Recursive inference to train
your own models and agents

Start on frontier open weight models through one endpoint. Every completed task becomes training signal for models you own.

Start free, $25 tokens Book a demo

builder · cloud agent

Fix the flaky checkout test and open a PR.

bun test42 pass · 1 flaky · checkout settles after assert

readsrc/routes/checkout.ts · settle() races the cart write

editawait settle(cart) before the redirect

bun test43 pass · 0 flaky

open pull requestfix: settle cart before checkout redirect

The race was in cart settlement. Tests are green and the pull request is ready for review.

pull requestopentests43 passsession4m 12s

one OpenAI-compatible endpointmetered per secondsingle-tenant cellsweights export any timeoutcomes become training data

Recursive

Rented models stay flat. Yours improves with use.

Every task your agents finish is captured, bound to the outcome that followed it, and folded into the next version of a model you own.

By the benchmarks

Task success on repeat production work, Allocate against the agent builders and frameworks most teams start with.

Task success rate (%)

05101520253035404550AllocateCopilot StudioBedrock AgentsAgent frameworksIn-house stack

Internal benchmark · repeat production tasks after 30 days

AllocateAgent builders

Who owns the trained weightsAllocateYou do, exportableAgent buildersThe vendor

Learns from your outcomesAllocateEvery completed taskAgent buildersNot at all

Runs on open weight modelsAllocateAny model in the catalogAgent buildersVendor's models only

Billed onAllocateWork delivered, per secondAgent buildersSeats and capacity

Where your data trainsAllocateInside your own cellAgent buildersShared infrastructure

Cloud agents

Every agent runs on its own computer in your cell, on schedules or on demand. Write paths stay human approved.

ops · scheduled agent

Reconcile yesterday's provider invoices and flag mismatches.

readinvoices/2026-08-10.csv · 50 accounts

queryledger entries · 50 accounts matched against invoice lines

draft2 mismatches · duplicate charge on partial-window retry

proposecredit note for review · human approves before send

Two accounts were double charged on retried windows. The credit notes are drafted and waiting for your approval.

accounts50mismatches2actionawaiting approval

Recursive harness

Agents recall lessons before every turn and refine what they learned after every outcome.

allo · recursive harness

import { createHarness } from "allo" const harness = createHarness() const lessons = await harness.recall(task) await harness.refine({ outcome: "merged" })