Skip to content

Fleets, brokers, and compute

Three concepts describe where a managed Nori session runs: the fleet, the broker, and the compute provider. None of them is the coding agent or model provider.

A fleet is an organization’s pool of session machines and the policy governing them. Fleet settings cover:

  • Desired pool size and lifecycle timeouts.
  • Default and enabled agents, models, and skillsets.
  • Workspace repositories and the organization setup script.
  • Slack access and agent-reach policy.
  • Default timezone and behavior for triggers.
  • Optional distribution of Nori and infrastructure credentials.

The broker continuously reconciles the actual pool with those settings.

The broker coordinates the organization. It receives requests from chat surfaces, the web dashboard, Handroll, and triggers; claims and prepares machines; starts agents; records sessions; distributes credentials; and returns events to the initiating surface.

Hosted customers receive a single-tenant broker at <org>.norisessions.com. Dedicated deployments can run the broker against organization-controlled infrastructure.

The compute provider supplies the machine. It is selected when a hosted fleet or dedicated deployment is configured, not for each session.

Backend Availability Important characteristics
Modal Preferred hosted backend for new fleets Managed sandbox runtime; supports a separate per-org VM Sandbox option in alpha
Sprites Supported for existing hosted fleets Legacy hosted backend with snapshot and platform wake capabilities
AWS EC2 Dedicated BYOC deployments One private EC2 instance per session, with Systems Manager and temporary S3 staging
Docker Local deployments Worker containers on a prepared image and private network
E2B Internal alpha only Gated to Nori internal environments and an explicit feature flag

The product source still has a sprites configuration fallback for hosted fleets without an applied provider setting. Operationally, new hosted fleet deployments prefer Modal. The dashboard should therefore be treated as the source for an individual organization’s active backend.

These are independent choices:

  • Compute provider: where the session machine runs.
  • Agent: which coding product runs inside the machine.
  • Model provider: which model API and credential the agent uses.

For example, Codex can run in a Modal-backed hosted fleet or an AWS EC2 BYOC deployment. Changing the compute backend does not automatically change the agent.

Every backend implements the same core machine contract for lifecycle, command execution, filesystem access, readiness, and the agent bridge. Optional capabilities differ:

  • Sprites supports snapshots and platform wake behavior.
  • E2B supports suspend and resume, but remains internal alpha.
  • Modal, AWS EC2, and Docker recycle machines instead of presenting the same snapshot or suspend behavior.
  • Modal can use a gVisor sandbox or a separately configured VM Sandbox runtime.

These differences affect provisioning and checkpoint strategy, not the basic session interface exposed to users.

A backend switch is a fleet or deployment operation. Hosted provider changes drain and rebuild the pool. AWS EC2 and Docker are deployment-only modes and do not write themselves into the hosted fleet provider setting.

For a concise availability table, see supported agents and compute.

this page as markdown: /sessions/concepts/fleets-and-providers.md· index for agents: /llms.txt