News · 2026-10-08

Docker Agent adds public-repository skills and evaluation plumbing in a new release

Docker released Docker Agent v1.149.0 on October 7, adding public-repository skill loading, a dedicated evaluator backend and routing updates to its agent runtime. The project lets developers define models, tools and delegation in configuration files. The news is an update to an existing product, with new ways to bring reusable procedures into an agent and evaluate its behavior.

Key facts

• Docker Agent v1.149.0 was released October 7.
• The release adds public-repository skills, evaluation plumbing and interface or tool-hook fixes.
• Agent definitions use YAML and can be shared through container-image registries.
• The primary release record is Docker’s GitHub release page.

The repository README describes a declarative builder and runtime for agents. A developer names the agents, selects models and gives each one tools and delegation relationships. Running that configuration starts the workflow. This puts the surrounding system in a reviewable file, rather than leaving every operational decision hidden inside a long chat session.

A useful analogy is a restaurant’s station plan. One worker takes orders, another prepares food and another checks what leaves the kitchen. The plan records who does which job and which equipment each person can use. In an agent configuration, the stations are model-driven workers and the equipment is tool access. The plan can organize the workflow, but it does not prove the workers make good decisions or the equipment is safe for every use.

The release is incremental. It includes a dedicated evaluation backend, routing changes and fixes around the terminal interface and tool hooks. Those additions may improve how teams measure and operate agents, but a changelog is not a benchmark result. The material reviewed establishes the features and release date, not a quantified reliability improvement across workloads.

Public-repository skill loading is the most consequential workflow change. A skill can package instructions and supporting resources for a repeated procedure. Loading it means an agent can receive a method without the user rewriting it into every prompt. That makes discovery, selection and version control part of agent design. The related story on skills becoming a package format describes the broader shift from isolated prompts toward reusable procedural assets.

The feature also introduces a practical trust question. A public procedure can be useful, irrelevant or misleading; accompanying code can carry its own permissions and dependencies. That observation is an architectural implication, not evidence that Docker’s release contains a malicious skill or a discovered vulnerability. Teams still need to understand what a loaded package instructs the agent to do and what powers the configured tools give it. The prompt-injection lesson explains how external instructions can redirect behavior when authority boundaries are weak.

Usefulness must be evaluated separately from provenance. The new SkillsBench study tested 87 tasks across nine model-and-harness configurations. Its authors found aggregate gains, yet the same skills helped some configurations and hurt others on 32 tasks. That study did not evaluate Docker’s specific release. It supplies a reason to measure a skill in the destination workflow rather than treating a package’s existence as evidence of benefit.

Docker Agent also supports different model providers, tool connections, memory and retrieval utilities. Configurations can be shared through container-image registries. Those capabilities make the project a way to assemble and distribute an agent system, not simply a model wrapper. The existing lesson on agent harnesses explains why the code around a model can substantially change what the combined system accomplishes.

The project’s history corrects a tempting launch narrative. In the Hacker News discussion, Docker participant dgageot says it began roughly a year and a half earlier as cagent. Another Docker participant, aheritier, describes the original idea as “Compose for agents.” That phrase is an apt account of the configuration-based approach, but the project has broadened beyond a simple Docker-only execution concept.

Reception includes confusion about how the runtime differs from orchestration frameworks and sandboxes. One practitioner complained about brittle session behavior in a Codex integration; a Docker engineer replied that the path was not the team’s most common internal use and invited details. Those exchanges reveal integration questions, not a representative quality verdict. They also show why a recognizable infrastructure brand should not substitute for task-level testing.

For developers, this is a usable open-source update with a public changelog and a concrete configuration surface. The honest caveat is that orchestration, isolation, skill provenance and measured success are separate properties. A new skill loader and evaluator backend help teams organize those questions; they do not answer them automatically.

Primary source, verified: read the paper →

Ground Truth. (2026, October 8). Docker Agent adds public-repository skills and evaluation plumbing in a new release. Ground Truth. https://groundtruth.day/news/docker-agent-public-repository-skills-release.html

@misc{groundtruth:docker-agent-public-repository-skills-release,
  title  = {Docker Agent adds public-repository skills and evaluation plumbing in a new release},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/docker-agent-public-repository-skills-release.html}
}

Topics: developer-tools · agents · open-source · agent-skills · orchestration