Key takeaways
- REST-based protocol using standard HTTP patterns — lower barrier to entry than JSON-RPC or gRPC-based alternatives
- Multimodal messaging — agents exchange text, code, files, and media natively, not just text
- Conversation-oriented — built around runs and sessions rather than task lifecycle states
- IBM-backed with BeeAI Platform integration and DeepLearning.AI course
FAQ
What is ACP?
ACP (Agent Communication Protocol) is an open protocol enabling AI agents, applications, and humans to communicate using multimodal messages over standard REST endpoints.
How does ACP differ from A2A?
ACP uses REST/HTTP conventions and focuses on conversation-oriented messaging. A2A uses JSON-RPC 2.0 and focuses on task lifecycle management. ACP is simpler; A2A has broader industry adoption.
Who created ACP?
ACP was created by IBM's BeeAI team (i-am-bee) and is open-source under Apache 2.0.
Is ACP production-ready?
ACP has production features like high availability support (Redis/PostgreSQL), distributed sessions, and trajectory metadata, but the ecosystem is smaller than A2A's.
Executive Summary
ACP (Agent Communication Protocol) is IBM's open protocol for agent-to-agent and agent-to-human communication, built on standard REST conventions. Where A2A uses JSON-RPC and task lifecycle management, ACP takes a simpler approach: agents communicate via multimodal messages over familiar HTTP endpoints. [1]
The protocol powers the BeeAI Platform — IBM's agent discovery and sharing platform — and has a growing ecosystem including JetBrains AI Assistant integration and a DeepLearning.AI course.
| Attribute | Value |
|---|---|
| Organization | IBM (BeeAI / i-am-bee) |
| Launched | 2025 |
| License | Apache 2.0 |
| GitHub Stars | ~950 |
| SDKs | Python, TypeScript |
Product Overview
ACP enables agents to communicate using rich, multimodal messages — not just text strings. [2]
Key Capabilities
| Capability | Description |
|---|---|
| Agent Manifest | Discovery document describing capabilities, name, and metadata |
| Runs | Single agent execution with inputs, supporting sync, async, and streaming |
| Multimodal Messages | Text, code, files, media in a single message structure |
| Sessions | Stateful conversation context across multiple runs |
| Distributed Sessions | URI-based session continuity across server instances |
| Trajectory Metadata | Track multi-step reasoning and tool calls across message parts |
Protocol Design
| Aspect | Detail |
|---|---|
| Transport | REST/HTTP(S) |
| Wire Format | JSON (OpenAPI-specified) |
| Discovery | Agent Manifests |
| Communication | Sync, streaming (SSE), async |
| State | Session-based (optional) |
| HA Support | Redis/PostgreSQL backends |
Strengths
- REST simplicity — Standard HTTP patterns integrate into existing infrastructure without specialized clients
- Multimodal native — Rich message types beyond text from day one
- IBM backing — Enterprise credibility and resources
- Distributed sessions — Session continuity across instances is a production-ready feature
- Trajectory metadata — Unique feature for tracing multi-step agent reasoning
- JetBrains integration — ACP agents available in JetBrains AI Assistant
- Clean OpenAPI spec — Machine-readable API definition enables code generation
Cautions
- Smaller ecosystem — ~950 stars vs A2A's 22K; fewer supporting organizations
- IBM-centric — BeeAI Platform is the primary deployment target
- Last push Aug 2025 — GitHub repo hasn't been updated in 6 months, raising questions about active development
- Fewer SDKs — Python and TypeScript only (A2A has 5 languages)
- No gRPC — REST-only limits high-throughput use cases
- Limited enterprise adoption signals — No named production customers outside IBM ecosystem
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Open Source | Free | Full protocol, SDKs (Apache 2.0) |
| BeeAI Platform | Free/TBD | Agent discovery, hosting, sharing |
Competitive Positioning
| Protocol | ACP Differentiator |
|---|---|
| A2A | ACP is REST-native and simpler; A2A has broader adoption and richer task lifecycle |
| ANP | ACP is centralized/enterprise; ANP is decentralized with DID-based identity |
| Summoner | ACP is conversation-oriented; Summoner focuses on cross-org durable transactions |
| MCP | MCP is agent-to-tool; ACP is agent-to-agent. Complementary |
Ideal Customer Profile
Best fit:
- IBM Cloud / BeeAI Platform users
- Teams wanting simple REST-based agent communication
- JetBrains ecosystem developers
- Organizations building conversation-oriented agent workflows
Poor fit:
- Teams needing broad multi-vendor interoperability (A2A has more adoption)
- High-throughput agent networks needing gRPC
- Decentralized/cross-organizational trust scenarios
- Teams needing SDKs beyond Python/TypeScript
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | Strong — IBM-backed |
| Market Position | Niche — Competing against A2A's momentum |
| Innovation Pace | Slowing — Last GitHub push August 2025 |
| Community/Ecosystem | Small — ~950 stars, limited adoption |
| Long-term Outlook | Uncertain — May converge with or be absorbed by A2A |
Bottom Line
ACP is a well-designed, REST-native protocol that's simpler to adopt than A2A. The multimodal messaging and trajectory metadata are genuinely useful features. But the ecosystem gap is significant: A2A has 20x the GitHub stars, 5x the SDK languages, and 150+ supporting organizations.
Recommended for: IBM/BeeAI ecosystem users and teams that prefer REST simplicity over JSON-RPC complexity.
Not recommended for: Teams building multi-vendor agent ecosystems — A2A's network effects make it the safer bet.
Outlook: ACP's best path forward is either A2A convergence (adopting A2A's wire format while keeping ACP's REST simplicity) or deep IBM enterprise integration. Standing alone against A2A's momentum is an uphill battle.
Research by Ry Walker Research • methodology