Skip to main content

Agent Runtime

The Agent Runtime is the execution environment that powers each user's AI workspace. It runs inside a dedicated, isolated pod provisioned for that user and handles all AI workload execution on their behalf.

What It Does

  • Runs as a private service inside a per-workspace pod
  • Manages a coordinated set of AI agents working together on behalf of the user
  • Handles multi-turn conversations with full context across agent transitions
  • Drains in-flight work gracefully on shutdown

Multi-Agent Execution

Each runtime hosts multiple specialized agents that coordinate to fulfill user requests. A routing layer determines which agent is best suited to handle a given task and delegates accordingly. Agents share context within a session and hand off work cleanly.

Tool Access

Agents have access to a curated set of platform tools — covering areas like web research, persistent memory, and user notifications — without holding raw API credentials. All tool use is brokered through the platform and recorded in an audit log.

Communication

The runtime communicates with the orchestrator over an authenticated, encrypted channel. There is no public endpoint. The runtime is reachable only over cluster-internal networking, and every request is verified before execution.

Deployment

Each user gets exactly one Agent Runtime pod in the platform's shared runtime namespace. Resource controls and backend mediation enforce isolation between users.

PropertyValue
NamespaceShared runtime namespace
NetworkingClusterIP service per pod
LifecycleManaged by the platform's runtime operator

See Runtime Lifecycle for the full state machine.

What's Next

See Runtime Lifecycle to understand how the platform provisions and manages the pod that runs the Agent Runtime.