Навык docker-agent-deploy

Docker Agent: Serving, Sharing, and Evaluating

Overview

This skill owns the integration surface of Docker Agent: making an agent reachable by other software (docker agent serve), distributing it through an OCI registry the way container images are distributed (docker agent share), and proving it still behaves after a change (docker agent eval). It assumes the agent config already exists — see docker-agent-config for authoring it, and docker-agent-run for interactive/local invocation.

When to use this skill

Activate this skill when:

• The user wants an agent reachable over MCP, an OpenAI-compatible chat endpoint, a plain HTTP API, or A2A/ACP.
• The user wants to publish an agent to Docker Hub (or any OCI registry) or pull one someone else published.
• The user wants automated evaluations (regression tests) for an agent, or wants to gate CI on eval results.

Do not use this skill when

Do not use this skill when:

• The task is authoring the agent.yaml itself (models, toolsets, sub_agents) — use docker-agent-config.
• The task is running the agent interactively on a developer's machine, choosing --safety/--sandbox, or aliases — use docker-agent-run.

Core guidance

Serving an agent

• Five server modes, each with its own default loopback listen address — never expose any of them beyond loopback without authentication:

  Mode	Default listen	Auth flag	Has --safety?
  serve mcp	127.0.0.1:8081	--auth-token (only with --http)	Yes (only with --http)
  serve api	127.0.0.1:8080	--auth-token	No
  serve chat	127.0.0.1:8083	--api-key / --api-key-env	Yes
  serve a2a	127.0.0.1:8082	--auth-token	Yes
  serve acp	(stdio only)	n/a	No

  docker agent serve mcp ./agent.yaml --http --listen 127.0.0.1:9090 --auth-token "$TOKEN"
  

• serve mcp defaults to stdio transport (for local clients like Claude Desktop); pass --http only when you need a network-reachable MCP endpoint, and set --auth-token whenever you do.

• Binding any server flag to a non-loopback address without an auth token/key is refused; --insecure-no-auth exists to force it and must be treated as a deliberate, documented exception, never a default.

• serve mcp (with --http), serve chat, and serve a2a expose --safety (strict/balanced/restricted/autonomous); Docker's docs state it defaults to restricted for these modes when unset. serve api and serve acp expose no --safety flag at all. Never raise --safety to autonomous on a network-reachable listener; if a served agent must approve more, prefer balanced and keep auth enabled.

• serve api accepts a directory instead of a single file: every .yaml/.yml/.hcl in it is exposed under /api/agents. Use --session-workingdir-root to confine session working directories when the server is reachable by more than one user.

Sharing agents via OCI registries

• Push and pull agent configs the same way you push and pull images — same registry, same docker login auth:

  docker agent share push ./agent.yaml docker.io/username/my-agent:latest
  docker agent share pull docker.io/username/my-agent:latest
  

• instruction_file contents are inlined into the pushed artifact automatically, so a published agent stays self-contained — you do not need to bundle the referenced files separately.
• Pin sub_agents that reference the pushed artifact to a digest (name@sha256:...) once published, to avoid a per-run registry lookup and to guarantee the exact config a consumer gets.
• Use --force on share pull only when you intend to overwrite a local copy that already exists; without it, an existing local config is left untouched.

Evaluating agents

• Evals live in an evals/ directory next to the agent config by default; each eval is one JSON session file capturing a user message, the recorded tool calls, and an evals object with the scoring criteria.
• Create eval sessions from real conversations rather than hand-writing JSON: run the agent interactively, then use the /eval slash command in the TUI to save the session, and edit in relevance/size/assertions criteria afterward.
• Four scoring dimensions: Tool Calls (F1 against the recorded sequence), Relevance (LLM-judge, --judge-model, default anthropic/claude-opus-5), Size (S/M/L/XL response-length bucket), and Assertions (deterministic checks; see the complete assertion-type list in references/eval-format.md). Prefer assertions over relevance when a check can be exact: they need no judge model and are deterministic, not approximation-prone.
• Evaluations run inside containers for isolation; a Docker-compatible runtime is required. Dedicated provider API keys (ANTHROPIC_API_KEY/OPENAI_API_KEY) are forwarded automatically. GITHUB_TOKEN/GH_TOKEN are not forwarded automatically (they're broad host credentials, not model keys) — pass them explicitly with -e GITHUB_TOKEN when an agent's provider needs one (e.g. github-copilot).
• Gate CI on regressions, not on absolute scores, with --baseline:

  docker agent eval ./agent.yaml --baseline results/2026-08-01-run.json --regression-tolerance 0.05
  

  A previously-passing eval that now fails always gates regardless of tolerance; cost changes are reported but never gate. A baseline or run with zero evaluations (e.g. an --only pattern matching nothing) is rejected rather than reported as passing.
• Use --keep-containers plus your runtime's exec to inspect a failed eval's container; the eval's .db session file holds the full conversation for offline debugging.

Verify

• After changing a served agent's config, re-run its evals with the same explicit --safety value used in the deployment before restarting the listener — this catches an approval-policy regression before it reaches traffic. If a rollout must be rolled back, restore the prior config and safety flag; never restore an unauthenticated listener as a rollback shortcut.

Related skills

• For writing or changing the underlying agent.yaml, use docker-agent-config.
• For local/interactive runs, safety-mode choice, and sandboxing, use docker-agent-run.

References

• references/eval-format.md — full eval session JSON schema and CLI flag table.
• references/sources.md — provenance of every rule in this skill.

Assets

• assets/eval-session-example.json — a minimal eval session file to copy and adapt.

Checks

• checks/verification.md — Verification runbook for serving, sharing, and evaluating an agent.