Docker ships docker-agent, a CLI plugin for declarative multi-agent AI systems

Docker has released docker-agent, a CLI plugin for building and running AI agents that are defined entirely in declarative YAML. The project is published in Docker's GitHub organization at docker/docker-agent and was featured on the Hacker News front page, signaling strong developer interest in the launch.

A docker subcommand for agents

Once installed, the plugin adds an agent subcommand to the Docker CLI. An agent is described in a YAML file that names the underlying model, a natural-language instruction, and one or more toolsets, such as a reference to an MCP server. Running docker agent run agent.yaml starts the agent, while docker agent new generates a configuration interactively. The README's example defines a root agent backed by openai/gpt-5-mini with a DuckDuckGo search toolset exposed through MCP.

The design leans on a multi-agent architecture: a root agent can delegate work to a team of specialized sub-agents, and delegation happens automatically based on the configuration rather than through hand-written orchestration code.

Tools, models and retrieval

According to the repository's README, docker-agent ships with several notable capabilities:

• Multi-agent orchestration with automatic task delegation between specialized agents.
• A tool ecosystem that combines built-in tools with any MCP server, whether local, remote, or Docker-based.
• Provider-agnostic model support covering OpenAI, Anthropic, Gemini, AWS Bedrock, Mistral, xAI and Docker Model Runner.
• Built-in reasoning tools for thinking, task tracking and memory.
• Pluggable retrieval-augmented generation with BM25, embeddings, hybrid search and reranking.

Docker Model Runner is worth highlighting because it allows agents to run against local models, removing the need for a cloud API key.

Packaging and installation

Distribution follows the pattern Docker established with container images. Agents can be pushed to any OCI registry — for example myorg/agent:tag — and then pulled and run anywhere with docker agent run, turning an agent into a versionable, shareable artifact.

Installation options include Docker Desktop 4.63 or newer, where the plugin is pre-installed, Homebrew via brew install docker-agent, and standalone binaries from GitHub Releases that can be symlinked into the Docker CLI plugins directory. Users must supply at least one provider API key, or rely on Docker Model Runner for local inference.

Dogfooding and telemetry

Docker says it uses docker-agent to develop docker-agent itself: the repository includes a golang_developer.yaml agent configuration for contributors. The tool also collects anonymous usage telemetry, with details documented in the repository.

Why it matters

Docker is applying its proven playbook — standardized packaging and one-command distribution — to AI agents. By expressing an agent as a YAML file that can be versioned in git, shared through existing OCI registries, and started with a familiar Docker command, the barrier to building multi-agent systems drops considerably compared with code-heavy frameworks.

The MCP support also matters: rather than inventing another tool-integration format, Docker plugs into an ecosystem that tool vendors and model providers are already converging on. And by shipping the plugin inside Docker Desktop, agent tooling lands where many developers already spend their day.

There are caveats. The project is new, using hosted model providers means managing API keys, and anonymous usage telemetry is collected, so teams evaluating it should read the documentation before adopting it broadly. Still, the launch signals a broader shift: agent development is moving from bespoke application code toward mainstream infrastructure tooling, and Docker clearly intends to own a piece of that transition.