Agents that get better in production.

One platform to build, test, evaluate, and run your agents. Every run is a trace, every trace is a test, every test is a reward. Post-train models you own on the work your agents already do.

Talk to us Read the customer stories

Built by a team that ran infrastructure for sandboxes, training, and inference at AWS, Google, Rubrik, and NVIDIA.

See it in ninety seconds.

One agent built, run, scored, and compared across models, in the console.

The oncell runtime

One agent runtime, and the whole lifecycle comes with it.

Every agent action crosses one boundary. That boundary is the trace, the test, the reward, and the audit trail at once.

agent.ts

import { Agent, tools } from "oncell";

const agent = new Agent("support", {
  instructions: "You are Acme's support agent.",
  model: "claude-sonnet",
  tools: [tools.memory, tools.files, crmLookup],
  budget: { perDay: "$5" },
});

agent.chat(({ message }) => agent.llm(message));

agent.eval({
  input: { message: "Refund my order" },
  rubric: "Polite, cites policy, never promises a refund",
});
agent.eval({
  input: { message: "Reset my password" },
  assert: (out) => out.cost < 0.01,
});

Five stages of running AI agents in production.

Agent evaluation, agent observability, and the agent runtime are sold as separate products. On oncell they are the same log, read five ways.

Buildone file

Instructions, model, tools, and a budget in one TypeScript file. Memory, a database, files, human approval, and schedules are built into the runtime.

Testfork

Declare what good looks like next to the agent. Fork any customer's agent into a staging copy in seconds and test against real state without touching it.

Evaluatescorecard

A scorecard per run, diffed against the last one: what got fixed, what regressed, how cost moved. Regressions block the release.

Runaudit

Deploy and let it work. The run log is the audit trail: every action, every approval, every dollar, per customer. Budgets are enforced below the model.

Improvereward

Run logs are trajectories and evals are the reward. Search prompts and models against it today; post-train models you own tomorrow. Then build again.

Improve

Every run is a trajectory. Every eval is a reward.

Post-training an agent needs an environment, a reward, and trajectories from real usage. oncell has all three as a by-product of running the agent.

Environment

A forked cell is a resettable copy of the agent with its real tools and state. Run a hundred rollouts, keep none of them.

Reward

The evals your team wrote to test the agent double as the objective: behavior, cost, and latency, scored the same way every time.

Trajectories

The run log already records every prompt, tool call, result, and approval. Export them, or let the platform search prompts and models against the reward.

How the loop works

Agents in production, with the cost per customer.

Teams running agents on oncell, with the number each one cares about: what an agent costs per customer, and the run that became a test.

All customer stories →

JustCopy.aiAI website builder · 30k users

The website was step one. The agents run the business.

JustCopy builds a small business its website, then the owner launches agents to run it: one that answers customers now, one that works in the background every day. oncell is the runtime for both, one agent per business, metered per business.

$0.0006

per customer conversation, metered per business

Read the story →

OpenCrewMulti-agent HQ · open source

One agent, one cell. The crew is the product.

OpenCrew is the HQ where AI agents work as a team. Every agent in a crew runs in its own oncell cell behind OpenCrew's HQ: its own tools, its own budget, its own run log, paused at no cost when idle.

cell per agent, with that agent's tools and nothing else

Read the story →

Multi-tenant AI agents, governed by the runtime, not the prompt.

The controls an enterprise needs are properties of where the agent runs, so they hold for every agent, every version, every customer.

Isolated by default

Every agent runs in its own cell with its own filesystem and no network except through the supervisor. One cell per customer, team, or tenant.

A complete audit trail

Every model call, tool call, approval, and dollar is recorded in the run log as a by-product of running. Nothing to instrument, nothing missed.

Spend enforced below the model

Daily budgets and per-run ceilings are enforced by the runtime, where a clever prompt cannot argue with them.

Humans in the loop

Agents park on approval for as long as it takes, at no cost, and resume exactly where they stopped. Every decision is logged.

Cost per customer

Model usage is metered at the boundary and attributed to the cell and the user, so you know what each agent costs each customer.

Your models, your data

Trajectories and evals stay yours and export in one command. The SDK is open source and runs locally. No shared models are trained on your work.

Single sign-on today. VPC deployment and SOC 2 Type II are part of the enterprise tier.

1 file

the agent and its evals, shipped together

100%

of agent actions in the audit log

while an agent waits on a human or a schedule

200 ms

to resume a parked agent from its snapshot

Questions teams ask before running agents in production

How is this different from a tracing tool like LangSmith or Braintrust?

They score what you instrument. oncell runs the agent, so it sees every call and its cost without instrumentation, can replay a run against a new version, and can fork production state to test against. Evals receive the full result, including what it cost, so a deploy can be gated on spend.

Do you train on our data?

No. Your run logs, evals, and trajectories are yours. The improvement loop optimizes your agents against your rewards; nothing is pooled across customers, and everything exports in one command.

What does “agents that get better” mean concretely?

Every production run becomes a candidate test case. Evals are the reward. Today the platform searches prompts and models against that reward and reports the trade-off in pass rate, cost, and latency. Next, the same trajectories and rewards feed post-training of open models you own, served through the same gateway.

Where does it run?

On oncell's cloud today, with SSO. VPC deployment and SOC 2 Type II are part of the enterprise tier. The SDK is open source and runs on a laptop for local development.

Are we locked in?

The SDK is open source, agents are plain TypeScript, and trajectories, evals, and reports export as files. What accumulates on oncell is the improvement history of your agents, which is why teams stay.

How is it priced?

Usage through the metered model gateway plus seats, with an enterprise agreement for VPC, SSO, and support. Talk to us for a quote against your agent count and traffic.

Bring one agent. Leave with a loop.

We will run your first agent with you, write its evals together, and show you what it costs per customer by the end of the week.

Book a 30-minute session Read the docs