# The Agentic OS Protocol

> The standard for multi-agent systems. Define agents, coordinate workflows, and build systems where agents work together.

## Quick Start
- [Introduction](/docs): Overview of the protocol
- [AGENT.md](/docs/agent): Define agents
- [SKILL.md](/docs/skill): Define skills

## Schemas

### Core
- [schema.json](/v1/schema.json): JSON Schema index

### Workflows
- [evaluator-optimizer.json](/v1/workflows/evaluator-optimizer.json)
- [orchestrator-workers.json](/v1/workflows/orchestrator-workers.json)
- [parallelization.json](/v1/workflows/parallelization.json)
- [routing.json](/v1/workflows/routing.json)

### Runs
- [approval.json](/v1/runs/approval.json)
- [cancel.json](/v1/runs/cancel.json)
- [retry.json](/v1/runs/retry.json)
- [run.json](/v1/runs/run.json)
- [timeout.json](/v1/runs/timeout.json)

## Docs
### protocol
- [Apps](/docs/apps): Distribution manifests for the Agentic OS — declare which vendors implement which protocol interfaces in a complete agentic system configuration.
- [Architecture](/docs/architecture): Architecture overview of the Agentic OS Protocol — six domains, the agent loop, workflow patterns, and how they compose into a system.
- [Introduction](/docs): Open specification for orchestrating AI agents at scale — standardized interfaces for system, context, actions, checks, workflows, and runs.
- [Motivation](/docs/motivation): Why the Agentic OS Protocol exists — the interoperability, coordination, and quality problems it solves for multi-agent systems.
- [MCP Servers](/docs/actions/mcp-servers): MCP Servers interface — discover, connect, and invoke tools from external Model Context Protocol servers within agent workflows.
- [System Actions](/docs/actions/system): System actions — aggregated write operations facade for filesystem, settings, and environment changes during the agent loop's action phase.
- [Tools](/docs/actions/tools): Unified tool discovery and execution for AI agents across MCP servers, built-ins, and custom providers.
- [Audit](/docs/checks/audit): Verification audit trail aligned with ISO 27001 and ISACA/ITAF standards.
- [Judge](/docs/checks/judge): LLM-as-judge evaluation — score agent output against quality criteria using configurable models, rubrics, and pass/fail thresholds.
- [Rules](/docs/checks/rules): Declarative verification criteria that agent output must satisfy before being accepted
- [Screenshot](/docs/checks/screenshot): Visual capture and baseline comparison for visual regression detection in agent checks.
- [Agent Loop](/docs/concepts/agent-loop): The agent loop — gather context, take actions, verify results. The core cognitive cycle that drives every AI agent in the protocol.
- [Agentic OS](/docs/concepts/agentic-os): The Agentic OS concept — an operating system metaphor where LLMs function as the kernel, managing agent resources through standardized APIs.
- [Agent Lifecycle](/docs/concepts/lifecycle): Agent lifecycle management — registration, discovery, execution, and evaluation. How the system manages agents from creation to retirement.
- [Workflows Taxonomy](/docs/concepts/workflows-taxonomy): Taxonomy of workflow patterns for AI agents — task, quality, and recovery workflows. Based on Anthropic's building blocks for agentic systems.
- [Embeddings](/docs/context/embeddings): Embeddings interface for AI agents — vector similarity search, semantic indexing, and namespace-scoped collections for RAG and retrieval.
- [Context](/docs/context): Context domain — read-only facades for the agent loop's gather phase. System state, embeddings, and key-value storage for AI agents.
- [Key-Value Store](/docs/context/kv): Key-value store for AI agents — flat, namespace-scoped persistence for structured data with TTL support and typed get/set/delete operations.
- [System Context](/docs/context/system): Read-only composition of all system state interfaces for the context phase of the agent loop
- [Approval](/docs/runs/approval): Human-in-the-loop approval gates — pause AI agent execution for review, with configurable timeouts, escalation, and multi-approver support.
- [Cancellation](/docs/runs/cancel): Workflow cancellation for AI agents — graceful abort with cleanup callbacks, partial result preservation, and cascading child run cancellation.
- [Retry](/docs/runs/retry): Retry mechanisms for AI agent workflows — configurable strategies with backoff, max attempts, and error filtering for failed operations.
- [Run Lifecycle](/docs/runs/run): Run lifecycle — states, transitions, and execution tracking for AI agent workflow runs. The core unit of observable work in the protocol.
- [Timeout](/docs/runs/timeout): Timeout control for AI agent workflows — set execution time limits with configurable strategies for graceful shutdown and cleanup.
- [Environment](/docs/system/env): Environment variable management for AI agents — read, write, list, and validate configuration in the host execution environment.
- [Filesystem](/docs/system/fs): Platform-level file system operations for reading, writing, and navigating files in the host execution environment
- [System](/docs/system): Infrastructure services that provide system-level intelligence for agent orchestration
- [Installer](/docs/system/installer): The kernel's package manager for installing skills, tools, and extensions at runtime.
- [MCP Client](/docs/system/mcp-client): Kernel-level interface for connecting to and managing external MCP servers that provide tools, resources, and prompts to the system.
- [Preferences](/docs/system/preferences): Per-agent and per-user preference storage for scoped configuration that customizes agent behavior without affecting the global system
- [Registry](/docs/system/registry): Registration and discovery of resources in the system — the kernel interface for managing what is available and finding it by criteria.
- [Sandbox](/docs/system/sandbox): Sandbox interface for AI agents — create isolated execution environments for running untrusted code with resource limits and lifecycle control.
- [Settings](/docs/system/settings): Global configuration that affects platform behavior across all agents and workflows.
- [Evaluator-Optimizer](/docs/workflows/evaluator-optimizer): Iterative generate-evaluate-optimize loop that refines output until quality criteria are met.
- [Workflows](/docs/workflows): Workflow patterns for AI agent coordination — routing, parallelization, orchestrator-workers, and evaluator-optimizer. Based on Anthropic's building blocks.
- [Orchestrator-Workers](/docs/workflows/orchestrator-worker): Break complex tasks into a plan, delegate steps to specialized workers, and synthesize results into a final output.
- [Parallelization](/docs/workflows/parallelization): Split a task into independent subtasks, execute them concurrently, and merge the results.
- [Routing](/docs/workflows/routing): Routing workflow — classify an input by intent or content type and delegate execution to a single specialized handler.