Agent Protocol — AI Glossary

Reference details

Related terms

• Agent-to-Agent Protocol (A2A)
• AI Agent
• Orchestrator

How it works

1. The protocol fixes a task-oriented REST surface: a call creates a task, further calls execute steps against it, and the task holds state between them. The task, not the message, is the addressable unit.
2. An agent implementing it exposes those endpoints regardless of what it is built from, so a harness can drive a LangChain agent, a bespoke loop or a vendor service through one interface.
3. Capability description lets a caller discover what an agent accepts before sending work, which is what makes substitution possible at runtime rather than at integration time.
4. Artifacts — files and structured outputs produced during execution — are retrievable against the task id, so results outlive the connection that requested them.
5. Step-level granularity is the feature that benchmarking depends on: a harness can advance an agent one step at a time and inspect what happened, rather than treating it as a black box that returns an answer.
6. What the protocol deliberately does not standardise is the agent's internals — its model, its tools, its prompting — which is why conformance says an agent is drivable and says nothing about whether it is any good.

Example: Why a shared interface changes the integration arithmetic

Suppose 10 agent implementations and 5 harnesses that need to drive them — benchmark runners, orchestrators, evaluation suites. Without a shared interface each pairing needs its own adapter: 10 x 5 = 50 adapters to write, and each must be maintained against changes on both sides. With a protocol, each agent implements it once and each harness speaks it once: 10 + 5 = 15 pieces of integration work, a reduction of 50 / 15 = 3.3x. The asymmetry grows with the ecosystem, since the adapter count is a product and the protocol count is a sum: at 50 agents and 20 harnesses it is 50 x 20 = 1,000 against 50 + 20 = 70, over 14x. That arithmetic is the entire argument for agent interoperability standards, and it also explains why they are hard to establish — the benefit accrues to the ecosystem while the cost of conforming falls on each implementer, who gets the full cost and a fraction of the gain.

Commonly confused with

Agent Protocol, A2A and MCP are three different boundaries and are routinely discussed as competitors. MCP connects a model to tools and data. A2A connects peer agents, potentially across organisations, with discovery and multi-turn negotiation. Agent Protocol standardises how a harness drives a single agent as a task-executing service, which is why its natural home is benchmarking and orchestration rather than agent-to-agent conversation. A system can use all three at once, and asking which to adopt usually means the real question is which boundary you are crossing.

When to use it

Reach for it when:

• Benchmarking agents, where a harness must drive many implementations without a custom adapter for each.
• Orchestration layers that should be able to swap the underlying agent implementation without rewriting the caller.
• Publishing an agent for others to drive, where conformance lowers the cost of someone adopting it.
• Step-level inspection, where advancing an agent one step at a time is what the evaluation requires.

Reach for something else when:

• Subagents inside one application, where an in-process call avoids a network boundary and a protocol for nothing.
• Peer-to-peer agent interaction with discovery and negotiation, which is what A2A addresses.
• Connecting a model to tools, which is MCP's boundary.
• As a quality signal. Conformance means an agent can be driven, not that it performs.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Agent Protocol — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path