Chinese
Uncategorized

Dependabot Can Update Agent Packages, but Review Still Matters

Dependabot can now update Agent Package Manager content. That brings agent instructions into a familiar review flow, with new risks.

Jacky Wang 4 min read 69 views

When a repository contains skills/, prompts, and instructions written for an agent, I treat them as part of the codebase. They may not compile, but they can change how code is edited, which commands are called, and what a reviewer is shown. On November 3, GitHub’s changelog announced Dependabot version updates support for Agent Package Manager. Dependabot can open pull requests for supported agent skills, prompts, and instructions.

The useful part is not the novelty of updating prompts automatically. It is the familiar review path: a diff, a pull request, a merge, and a rollback target. When agent rules are copied by hand, repositories drift quietly. The first symptom may not be a failure. It may be two agents making different assumptions about the same task.

Treat agent content as supply-chain input

I would start by inventorying the files that can influence an agent:

find . -maxdepth 3 -type f \
  \( -path './.github/*' -o -path './skills/*' -o -name 'AGENTS.md' -o -name 'CLAUDE.md' \) \
  -print | sort

That is an inventory command, not a recommendation to give Dependabot every matching path. I still want to verify where a package comes from, who maintains it, what it changes, and how it is invoked. A skill may only check formatting, or it may carry scripts that access the network and write files. The word “instructions” does not mean “no execution consequence.”

For a production repository, I would constrain automatic updates to a known directory and require the pull request to include a change summary. Major-version updates should remain manual. Any change involving commands, permissions, or external connectors should stop at review. Agent policy updates should not automatically share the exact merge gate used for a small npm patch.

A pull request is evidence of change, not proof of safety

Dependabot pull requests make the diff visible and give the team a clear rollback object. That is valuable. But the PR proves that files changed; it does not prove that a new skill has not expanded the agent’s blast radius.

During review, I would ask four concrete questions. What can the agent read? Which tools can it call? What happens when a command fails? Can it retry, continue, or silently fall back? Finally, can the task carry secrets, user data, or deployment authority into an external service?

Tests should cover those boundaries too. A regression test for a skill should not only assert that the happy-path task completed. It should check behavior when permissions are missing, the network is unavailable, or the input is malformed. For example, I would run a read-only investigation skill in a temporary container and verify that it cannot read a production kubeconfig or upload logs to an unapproved endpoint.

Keep repository version separate from runtime version

I would not say that a production agent uses a new policy just because a Dependabot PR merged. Repository content, the packaged bundle, the running agent, and the effective permissions are four separate states. Each needs its own verification.

A tiny but useful runtime record could look like this:

agent_policy=2026-11-07.1
skill_bundle_sha=<commit-sha>
mode=read-only

When an automated change looks strange, this lets me answer which policy was actually running rather than pointing at the current branch. If the policy can deploy software or spend money through an API, I also want an approver and an expiry time in the change record.

This release makes agent content behave more like a dependency, but I would resist treating it exactly like one. A software package is usually constrained by a version and tests. Prompts, skills, and instructions also change the decision path. My preference is to let Dependabot discover and propose updates while a human decides whether the new capability is worth the new permission surface.

The practical payoff for me is operational handover. A reviewer should be able to answer who approved the update, which repository or fixture tested it, and how to restore the previous bundle within a few minutes. Agent knowledge can move faster without making ownership vague. That is the standard I would want before allowing these pull requests to merge automatically.

I would also keep the update cadence boring. A weekly review window is easier to staff than a stream of instant merges, especially when the same skill is shared by several repositories. The goal is not to freeze agent behavior; it is to make a behavior change visible while someone still remembers why the old rule existed.

Primary references: GitHub’s November 2026 changelog and GitHub’s agent apps announcement.

发表评论