Core concepts

AgentMuru separates application intent from provider and transport details. The same agent can run with a deterministic fake, an official hosted-model adapter, in-memory state, SQLite, the Python API, or the browser Workspace.

Application model

An Application names one primary Agent, optional handoff targets, session and artifact stores, and user-facing metadata. The Runtime executes that definition and owns policy, approvals, tracing, cancellation, and event production.

Application
  Agent
    ModelProvider
    Tool[]
    permissions
  SessionStore
  ArtifactStore

Runtime
  PermissionPolicy
  ApprovalService
  Tracer

Read by concern

Concern	Concept
Define behavior and choose a model	Agents and models
Connect typed Python functions safely	Tools and approvals
Understand execution order and replay	Runtime and events
Keep conversations and runs	Sessions and memory
Store outputs and inspect execution	Artifacts and traces
Coordinate multi-step behavior	Workflows and handoffs

Design rule

Provider output is untrusted input. A model can request a registered tool, but it cannot grant permissions, approve itself, bypass argument validation, or write directly to a store. The runtime remains the authority for every side effect.