The Agent Connect Protocol (ACP) is a standardized API specification for enabling AI agent workflows, communication, and execution. It defines how AI agents can be configured, run, and interact with each other through a common interface. ACP is one element of a series of protocols introduced under the AGNTCY umbrella.
The protocol revolves around several key concepts:
- Agents: AI workflows that can be configured and executed. Each agent has capabilities, input/output schemas, and supported features.
- Threads: Session-like containers that group runs and maintain state between them. Threads track history and provide context continuity.
- Runs: Actual executions of an agent, which can be stateless (one-time executions) or part of a thread (maintaining history and context).
- Streaming: Real-time output capabilities allowing agents to provide incremental results.
- Interrupts: Mechanisms for agents to pause execution and request additional input before continuing.
Entity Relationship Diagram
Here’s a quick view of the primary entities.
Core Concepts
Agents
Agents are AI workflows that can be configured and executed. Each agent:
- Has a unique identifier
- Provides capabilities (what features it supports)
- Defines schemas for input, output, configuration, and state
- May support different protocol features (streaming, interrupts, etc.)