Levain LabsLevain Labs
Platform

How Agents Run on Levain

The six primitives behind every agent.

An agent is a program that reasons, plans, and acts. Running one reliably in production requires more than a prompt: it needs durable state, scheduling, isolation, and bounded execution. Levain provides six primitives that cover these concerns.

The primitives

PrimitiveWhat it gives you
SessionA durable log of everything the agent did.
OrchestrationScheduling, retries, and cancellation you can rely on.
HarnessStep-by-step execution inside the bounds you set.
SandboxA clean environment for every run, isolated from every other.
ResourcesThe data the agent can reach.
ToolsThe capabilities the agent can call.

Every feature in the product maps back to one of these six. When you read the API reference, the endpoints name them directly: a run is a single trip through the harness inside a sandbox, logged to a session. When you inspect a failed agent, you're reading its session. When you swap a data source, you're updating a resource.

Why these six

Each primitive addresses a distinct class of problem agents run into:

  • Unpredictable reasoning — the session records exactly what the agent considered.
  • Transient infrastructure failures — orchestration retries idempotently.
  • Runaway loops or cost — the harness enforces budgets and step limits.
  • Escaped boundaries — the sandbox isolates the run from other runs and the host.
  • Stale or wrong context — update the resource to point at fresh data.
  • Missing capability — register a new tool.

The primitives are independent. You can change the data a tool reads without touching the agent's instructions, tighten the harness budget without redeploying, or change sandbox backends without modifying the agent.

What you control, what we handle

YouLevain
The agent's goal and instructions.Keeping the agent on-task, on-budget, and inside safe bounds.
The data the agent needs.Making your data available to the agent on every run.
The capabilities the agent uses.Giving the agent the capabilities it needs, reliably.
The model provider (your key, or ours).Routing LLM traffic, tracking spend, and keeping costs honest.
When a run should start.Running it when you ask, and keeping it running through breaks.

Read the pages linked above in any order. The primitives are independent by design.

On this page