Docker Agent

Docker’s open-source runtime for building, running and sharing AI agents. Agents are defined in a declarative YAML file, can use built-in tools and any MCP server, and can delegate to each other as a team. It runs as a docker CLI plugin (docker agent) or as a standalone binary. Repository: docker/docker-agent (Apache-2.0).

A2A is not new

Docker Agent’s A2A support is not a fresh release. It has been in the project’s releases since v1.9.11 (2025-11-12, “Add a2a support”). The newest A2A-related change read is v1.150.0 (2026-10-09), which is a new release, not the first A2A release. The docs still label A2A as early support.

What it does (README and docs, read 2026-10-10)

• Declarative agents: an agent is a model, an instruction and a list of toolsets in YAML. Sample: docker agent run agent.yaml.
• Tools: built-in think, todo and memory tools; any MCP server (local, remote or Docker-based, including the docker: catalog references); RAG with BM25, embeddings, hybrid search and reranking.
• Multi-agent: teams of specialised agents that delegate tasks. The A2A tool lets an agent call a remote A2A agent over the network, as distinct from a local handoff.
• Models: OpenAI, Anthropic, Gemini, AWS Bedrock, Mistral, xAI and Docker Model Runner for local models.
• Sharing: agents can be pushed to and pulled from any OCI registry.
• Install: included in Docker Desktop 4.63 and later; Homebrew; or binary releases.

A2A support (docs and release notes)

• Serve: docker agent serve a2a ./agent.yaml starts an A2A server, by default on 127.0.0.1:8082. Non-loopback listeners need --auth-token unless --insecure-no-auth is passed.
• Agent card: the server publishes a standard A2A agent card. The docs don’t list its fields.
• Known limits, as listed in the docs: tool calls are handled internally and not exposed as separate A2A events; artifact support is not integrated yet; memory is not integrated yet; sub-agent scenarios “need further work”.
• Status: the docs carry an “Early support” warning, describe A2A as “functional but still evolving”, and give no dates or GA designation.
• Fixes in releases: a fix for the A2A tool silently succeeding on an empty sub-agent stream (v1.115.0), and a cap on oversized A2A tool results (v1.89.0).

Adoption and activity (snapshot 2026-10-10)

• Repository created 2025-09-01; about 4,400 stars; last push 2026-10-10; not archived.
• Releases run to v1.150.0 (2026-10-09), with several releases a week in the most recent weeks.
• The earlier releases link to docker/cagent pull requests, which suggests the project was renamed from cagent. This is an inference from the links, not a stated fact.

Context

Docker’s MCP tooling is covered in docker-mcp-gateway. For the protocol itself, see a2a-protocol. Other agent frameworks are in ai-agent-platforms-comparison.

Open items

• Which A2A specification version Docker Agent implements is not stated in the docs read.
• The agent card fields are not documented.
• The docs give no GA date. A2A support is labelled early, so treat it as experimental.
• No independent benchmark or security review was found.

Sources (read 2026-10-10)

• https://github.com/docker/docker-agent (README, releases via the GitHub API)
• https://docs.docker.com/ai/docker-agent/features/a2a/ (A2A features, serve command, limitations)