Article · 2026-07-18 · 6 moments

MCP vs A2A vs ACP: How AI Agents Actually Talk to Each Other

Agents are capable on their own. Combined with tools and other agents, their capabilities compound. ✦ AI generated

Mechanism

A2A lets an agent that can't complete a task alone discover a capable peer via a published Agent Card, delegate the task to it, and receive a structured result back, with the peer able to pause in an input-required state if it needs more information.

A2A enables one agent to find, delegate to, and exchange structured results with another agent via a published Agent Card.

transcript

ByteByteGo: An agent that can't complete a task alone discovers a capable peer via its Agent Card (published at a well-known URL), delegates the task, and receives a structured result back. If the second agent needs more input mid-task, it pauses in the input-required state and loops back to the first.

extends · 1gives example · 1

Claim◆

ACP took a REST-first approach with peers discovered via an Agent Manifest and called over HTTP before merging into A2A, and in production MCP and A2A are complementary rather than competing: MCP handles tool access while A2A handles agent-to-agent communication.

ACP's REST-based agent communication approach was folded into A2A, and today MCP and A2A are used together for tool access versus agent-to-agent coordination.

transcript

ByteByteGo: ACP took a REST-first approach. Peers were discovered through an Agent Manifest, called directly over HTTP, and responded to sync for low-latency tasks or via async SSE stream. In production, MCP and A2A are complementary. MCP handles tool access, A2A handles agent communication.

Claim

Although LLMs, RAG pipelines, and agents are different systems, evaluating any of them follows the same three-step recipe of picking a task, collecting eval data, and developing a grader, and every additional pipeline component becomes a new place for failures that evals must catch.

Evaluating LLMs, RAG pipelines, and agents differs in what's graded (retrieval, final answer, unit tests, coordination) but follows one shared underlying recipe.

transcript

ByteByteGo: LLMs, RAG pipelines, and agents are different systems, but the recipe for evaluating them is the same: pick a task, collect eval data, develop a grader. Every new component in the pipeline is a new place for things to go wrong, and a new thing your evals need to catch.

Data

Among current frontier open models, Inkling from Thinking Machines is now the strongest American open model with text, image, and audio input, Nemotron 3 Ultra from NVIDIA is a solid pick for long-running agents due to its cheap Mamba-hybrid long-context inference, and GLM-5.2 from Z.ai is currently the best open model for coding.

A roundup ranks Inkling as the top American open model, Nemotron 3 Ultra as best for long-running agents, and GLM-5.2 as the best open coding model, among other picks.

transcript

ByteByteGo: Inkling (Thinking Machines): released this week, now the strongest American open model. Text, image, and audio input. Nemotron 3 Ultra (NVIDIA): a solid choice for long-running agents. The Mamba hybrid keeps long-context inference cheap. GLM-5.2 (Z.ai): currently the best open model for coding.

extends · 1supports · 1

Mechanism◆

MCP lets a host application route a user request through an embedded MCP client to the correct MCP server, which executes the tool call and returns a structured response the agent uses to continue reasoning.

MCP standardizes how an agent's host application discovers and calls tools through an MCP client-server exchange.

transcript

ByteByteGo: The host app receives the user request, its embedded MCP client formats and routes it to the right MCP server, the server executes the tool call and returns a structured response. The agent uses the result to continue reasoning.

explains mechanism · 1provides context · 3

Mechanism

Redis's AOF persistence executes each command in memory first and appends it to an on-disk log afterward, which keeps writes non-blocking while still allowing the full dataset to be rebuilt by replaying the log after a crash.

Redis's Append-Only File mechanism logs commands after executing them in memory, enabling non-blocking writes and full dataset recovery via replay.

transcript

ByteByteGo: When a client sends a command, Redis first executes it in memory (RAM). After that, Redis logs the command by appending it to an AOF file on disk. This ensures every operation can be replayed later to rebuild the dataset. Since the command is executed first and logged afterward, writes are non-blocking.

Highlight slides

MCP: How a Tool Call Flows✦ from: MCP lets a host application route a user request through an embedded MCP client to the correct MCP server, which executes the tool call and returns a structured response the agent uses to continue reasoning.ACP: REST-First Agent Communication✦ from: ACP took a REST-first approach with peers discovered via an Agent Manifest and called over HTTP before merging into A2A, and in production MCP and A2A are complementary rather than competing: MCP handles tool access while A2A handles agent-to-agent communication.Closing the Loop✦ from: MCP lets a host application route a user request through an embedded MCP client to the correct MCP server, which executes the tool call and returns a structured response the agent uses to continue reasoning.ACP Merged Into A2A✦ from: ACP took a REST-first approach with peers discovered via an Agent Manifest and called over HTTP before merging into A2A, and in production MCP and A2A are complementary rather than competing: MCP handles tool access while A2A handles agent-to-agent communication.Production Roles: MCP vs A2A✦ from: ACP took a REST-first approach with peers discovered via an Agent Manifest and called over HTTP before merging into A2A, and in production MCP and A2A are complementary rather than competing: MCP handles tool access while A2A handles agent-to-agent communication.

Related episodes

Video · a16z Podcast

Software in the Age of Agents | The a16z Show

→

Video · a16z Podcast

Kavak's Playbook for Rebuilding a Company Around AI

→

Article · Simon Willison's Weblog

Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)

→