Docker Is Now Shipping AI Agents: docker-agent Turns a YAML File and an OCI Registry Into an Agent Distribution Channel

Docker's argument for how AI agents should be distributed looks a lot like its argument for how software should be distributed. The company has released docker-agent, an open-source CLI plugin that lets developers define AI agents in a declarative YAML file and run them with the same command grammar they already use for containers: docker agent run, docker agent new, and a registry pull that looks exactly like pulling an image.

The project, formerly known as cagent and now sitting at roughly 3,578 GitHub stars, is Apache-2.0 licensed and ships pre-installed inside Docker Desktop 4.63 and later. It is also available through Homebrew and GitHub Releases. The current build is v1.149.0, and the release cadence has been fast: v1.147 and v1.148 landed on Sunday and Monday, adding hook-driven agent routing, a 50 KiB bound on tool results so oversized output cannot stall compaction, and codemode tool-call visibility.

The model lineup is deliberately provider-agnostic. The README lists OpenAI, Anthropic, Google Gemini, AWS Bedrock, Mistral and xAI, plus Docker Model Runner for local inference, with a single environment variable switching between them. Tool access flows through the Model Context Protocol, supporting local, remote and Docker-based MCP servers, alongside built-in think, todo and memory tools. Retrieval-augmented generation is handled with BM25, embeddings, hybrid search and reranking.

The structural pitch is multi-agent orchestration. Instead of one monolithic agent, a root agent spins up sub-agents for specific tasks, each with its own model, toolset and instructions. Because the configuration is YAML, agent definitions can be versioned, reviewed and shared like infrastructure-as-code artifacts, and they can be pushed to any OCI registry and pulled anywhere with docker agent run myorg/agent:tag. Docker's own team runs docker agent run ./golang_developer.yaml as part of its development workflow.

The most recent release adds two lines worth noting. Skills entries can now point at a public GitHub repository URL and the runtime loads them, which puts Docker on the agent-skills standard in the same week that skill repositories dominated GitHub trending. And evaluators now route through the models gateway by default, with OpenAI Decisions available as a native evaluator backend alongside a dedicated evaluator service as an alternative to LLM-as-judge. Docker is among the first harnesses to wire a decision-model endpoint into the grading path.

Reception has been mixed. The project drew 126 points on Hacker News on Wednesday, with the familiar objection that the field already has LangChain, AutoGPT and CrewAI, and that YAML adds a layer of indirection between developer and model. Others noted that debugging distributed agent teams remains an unsolved problem regardless of framework. Docker also discloses anonymous usage telemetry, a standard but non-trivial detail given its enterprise customer base.

The strategic logic is still the strongest part of the story. Docker is bidding to make the container registry the distribution layer for agents, not just images. For organizations that already push images to registries every day, pushing an agent costs nothing new, and that distribution advantage is exactly what standalone agent runtimes do not have.