Discovering Docker Agent with AI Endpoints

Stéphane Philippart02/07/20264 min read

What if you could build a team of AI agents, each specialized in a different task, working
together like a well-oiled pipeline, without having to build agent orchestration logic from
scratch ?
That's exactly what Docker Agent brings to the table.

Docker Agent is an open-source framework for building collaborative multi-agent workflows.
Instead of one monolithic agent trying to do everything, you compose a team of specialized
agents. Each agent has a clear role, specific tools, and its own model, allowing them to
communicate and collaborate on complex, multi-step tasks.

In this post, we'll explore how to wire up Docker Agent with OVHcloud AI Endpoints to create
a multi-agent coding assistant: one agent browses documentation, another writes the code.

Install/configure Docker Agent

All detailed instructions for installing and configuring Docker Agent can be found in
the official documentation.
Once installed, Docker Agent is available as the docker-agent or docker agent command line
tool.
⚠ Docker Agent does not need Docker engine to run. ⚠

Authenticate with AI Endpoints

To use Docker Agent with AI Endpoints, export your API key:

bash

export OVH_AI_ENDPOINTS_ACCESS_TOKEN="your_api_key_here"

ℹ️ Docker Agent uses the token_key field in the provider configuration to know which environment variable holds the API token, here we name it OVH_AI_ENDPOINTS_ACCESS_TOKEN. ℹ️

Configure the provider and models

AI Endpoints is fully integrated in docker agents. That's mean that the configuration is very easy as you can see in the official documentation.

One of the strengths of AI Endpoints is the ability to choose the right model for each task. This allows each agent to use a model optimized for its specific role rather than forcing a single model to handle every step of the workflow. Here we use gpt-oss-120b for the crawler, which handles documentation analysis, web browsing and reasoning, while Qwen3.5-397B-A17B is used for the coder due to its strong code generation capabilities.

Create a team of agents for coding

Let's build a two-agent team: a crawler that fetches documentation from the web, and a coder that turns that documentation into working code.

The crawler uses the fetch toolset to read URLs and passes the extracted context to the coder via the sub_agents mechanism. The coder uses the filesystem toolset to create files and the shell toolset to execute or validate them when needed.

yaml

agents:
  crawler: 
    model: ovhcloud/gpt-oss-120b
    description: Documentation crawler
    instruction: |
      You are a documentation crawler.
      Step 1: use the fetch tool to read the URLs provided by the user and     extract the relevant information.
      Step 2: you MUST invoke the coder sub-agent tool with the full context you gathered, so it can write the code.
      Do NOT write the code yourself. Always invoke the coder agent.
    sub_agents: [coder]
    toolsets:
     - type: filesystem
     - type: fetch

coder:
  model: ovhcloud/Qwen3.5-397B-A17B
  description: Coder
  instruction: |
    Write code based on the context provided.
    Ensure code quality and maintainability.
    Save files to disk using the filesystem toolset, use the current path.
   toolsets:
    - type: filesystem
    - type: fetch
    - type: shell

As you can see using an OVHcloud model is very easy: get the model name from the catalog and add ovhcloud/ to the name in the configuration file.

Run the agents

Once your configuration file is ready, run the team with:

bash

docker agent run ovhcloud-agent.yml

Here is the Docker Agent in action:

[Video: https://blog.ovhcloud.com/wp-content/uploads/2026/07/docker-agents.mov]

Conclusion

In this post, we saw how Docker Agent makes it easy to build collaborative multi-agent teams (without requiring Docker engine) and how AI Endpoints fits in naturally thanks to its full OpenAI API compatibility.

AI Endpoints gives you the flexibility to choose the right model for each agent in a multi-agent workflow, all from a single provider configuration. Rather than relying on one model for every task, you can combine specialized models for research, reasoning, coding, content generation, or data processing.

The crawler/coder example shown here is just one illustration of this approach. The same multi-agent pattern can be applied to a wide range of workflows wherein specialized agents collaborate to complete complex tasks more efficiently.

Have questions or feedback? Join the dedicated #ai-endpoints channel on our Discord server.

To learn more about AI Endpoints, please explore our previous blog posts.

Share on: