Skip to content

Runtime Overview

The runtime is the execution plane for Agirunner.

It is the system that takes a claimed task and turns it into actual work:

  • claim the task
  • prepare the workspace
  • expose the allowed tools
  • run the agent loop
  • capture results and artifacts
  • report status, logs, and outputs back to the platform

An agent platform can route work without actually being able to perform it safely. The runtime exists so Agirunner has a real execution boundary instead of a thin wrapper around prompt calls.

That boundary matters because the runtime is where you enforce:

  • isolated workspaces and execution containers
  • tool transport and validation
  • execution budgets and loop guardrails
  • result capture and reporting contracts
  • execution-layer security and observability

The runtime currently runs two task-loop styles:

  • Reactive for specialist execution where fast tool use matters more than explicit planning
  • TPAOV for orchestrator execution where the system benefits from a structured think, plan, act, observe, verify cycle

The platform selects the mode through the task contract. The runtime enforces it.

  • task claiming and heartbeats
  • built-in tools and dynamic MCP tools
  • workspace materialization and repository handling
  • execution containers and image customization
  • structured capture of outputs, logs, and artifacts
  • health, metrics, traces, and execution logs