Docker Agent Now Loads Skills From GitHub and Uses OpenAI Decisions as the Judge
Docker's agent runtime, the docker agent CLI plugin that ships pre-installed in Docker Desktop 4.63+, hit Hacker News on Wednesday with 126 points on the same day it cut release v1.149.0. The project is the former cagent, now at 3,578 stars, and the idea has not changed: define agents in YAML, give them MCP tools, compose teams that delegate, push the whole thing to an OCI registry and pull it anywhere. Providers are OpenAI, Anthropic, Gemini, Bedrock, Mistral, xAI and Docker Model Runner for local models.
Two lines in Wednesday's release notes are the reason to look again. First, skills entries can now point at a public GitHub repository URL and the runtime loads them, which puts Docker on the agent-skills standard the same week GitHub trending is dominated by skill repos. Second, evaluators now route through the models gateway by default and OpenAI Decisions is a native evaluator backend, alongside a dedicated evaluator service as an alternative to LLM-as-judge. Docker is one of the first harnesses to wire the decision-model endpoint into the grading path, and it did so within a day of the beta going public.
The rest is the steady cadence this project has kept: v1.147 and v1.148 on Sunday and Monday added hook-driven agent routing, a 50 KiB bound on tool results so oversized output cannot stall compaction, and codemode tool-call visibility; v1.145 added W3C traceparent propagation across the 13 ACP methods. Pricing now follows the OpenAI service tier actually served rather than the one requested, which is the kind of billing honesty that only matters once you run thousands of agent turns a day.
Docker's bet is that agents are a packaging problem, which is the problem Docker already solved once. Brew install docker-agent, or run docker agent new.
Link: github.com/docker/docker-agent
← Back to all articles
Two lines in Wednesday's release notes are the reason to look again. First, skills entries can now point at a public GitHub repository URL and the runtime loads them, which puts Docker on the agent-skills standard the same week GitHub trending is dominated by skill repos. Second, evaluators now route through the models gateway by default and OpenAI Decisions is a native evaluator backend, alongside a dedicated evaluator service as an alternative to LLM-as-judge. Docker is one of the first harnesses to wire the decision-model endpoint into the grading path, and it did so within a day of the beta going public.
The rest is the steady cadence this project has kept: v1.147 and v1.148 on Sunday and Monday added hook-driven agent routing, a 50 KiB bound on tool results so oversized output cannot stall compaction, and codemode tool-call visibility; v1.145 added W3C traceparent propagation across the 13 ACP methods. Pricing now follows the OpenAI service tier actually served rather than the one requested, which is the kind of billing honesty that only matters once you run thousands of agent turns a day.
Docker's bet is that agents are a packaging problem, which is the problem Docker already solved once. Brew install docker-agent, or run docker agent new.
Link: github.com/docker/docker-agent
Comments