Google Labs' stitch-skills packages AI agent workflows so several coding agents can reuse them
Google Labs released a library of reusable agent workflows that runs across several coding agents, but it leaves retries, tracing and security undefined.
The short version
- stitch-skills implements the Agent Skills open standard and sits one layer above MCP servers.
- The same skill definitions work in Codex, Claude Code, Cursor, Antigravity and Gemini CLI; OpenCode needs manual installation.
- Skills are shared through Git, with no central registry and no way to update every skill at once.
- The standard does not define retries, resuming after a crash, tracing or limits on which tools a skill can call.
- The dev.to author calls it a reference implementation, not a production-ready framework.
Why it matters
If you build or configure AI agents, a workflow you write once can now travel between tools instead of being rewritten for each one. The catch is that the standard is young, and it says little about what happens when a step fails or who is allowed to call what. For non-engineers, the lesson is simpler: a good AI routine is worth saving as a reusable asset.
What Google Labs released
Google Labs has published stitch-skills, a library built on an open standard called Agent Skills. As dev.to reported in a technical write-up, it sits one layer above MCP servers. MCP, the Model Context Protocol, is a common plug that lets AI agents connect to outside tools.
The repository has 8.7K stars, a rough measure of developer interest. The same skill definitions work in Codex, Claude Code, Cursor, Antigravity and Gemini CLI. A sixth tool, OpenCode, works too, but only if you install the skills by hand.
The article stresses that this is not another MCP server. It is a way of distributing workflows that use MCP tools, so the interesting part is the packaging rather than any single tool.
Tools are verbs, skills are recipes
The write-up draws a clear line. An MCP tool is a single action, such as reading a file, running a command or searching the web. A skill is a multi-step recipe that chains several of those actions together, with instructions, expectations about what information it needs, and checks on the result.
According to the article, a skill file carries a few standard pieces:
- Descriptive details such as name, description, version, author and tags.
- Instructions for the agent, including places to insert user input.
- A list of the MCP tools the skill depends on.
- A statement of what the skill expects to receive and what it returns.
Stitch Skills uses this for jobs like generate-design, which calls design tools, checks the output and formats the result, and scaffold-component, which reads templates, generates code and writes files.
When an agent loads a skill, it reads the details, confirms the required tools are available, adds the instructions to its working context and offers the skill as something the user can call. Nothing runs until the user asks. Then the agent follows the steps in order and keeps track of results between calls.
Shared through Git, installed differently everywhere
Distribution runs on Git rather than a package store. In Codex, a command adds the repository as a marketplace and uses sparse checkout, a Git feature that downloads only chosen folders. That way an agent fetches just the skill groups it needs. The repository splits skills into design, build and utility plugins.
This keeps installs small, and you can pin a branch, tag or commit. The cost is that there is no central registry, no automatic handling of conflicting dependencies and no built-in rollback. If you install the design plugin today and the build plugin tomorrow, they may point to different versions, and you cannot update everything in one step. The article's suggested workaround is to fork the repository and pin your own version, which adds upkeep.
Each agent also has its own installation habits. Codex, Claude Code and Cursor use different commands, folders or scopes. OpenCode has no support for the install tooling and requires copying files. It is also stricter about naming: the skill's name must match its folder name exactly, so renaming a folder can quietly break a skill that worked elsewhere.
Where the portability story breaks down
The most useful part of the write-up is its list of things the standard leaves open. It does not say whether a skill should retry a failed tool call or give up. It does not say whether an agent can resume a skill after a crash. It offers no common way to trace calls across skills. And it does not stop a skill from calling tools beyond those it declared.
Each agent fills those gaps its own way. The article says Codex retries failed tool calls up to three times, Claude Code stops at the first failure, and Cursor has no documented retry behavior. So the same skill can behave differently depending on where it runs.
On security, the article's point is that a skill lists its tool dependencies up front, but nothing enforces that list at runtime. Agents simply trust skills to behave. Debugging a workflow that spans several skills is hard without agent-specific logs.
Who should use it, and how far to trust it
The author recommends the approach for teams that use several coding agents and want to share multi-step workflows, and who can live with Git-based versioning and manual dependency management. They advise against it where atomic versioning, tracing, security boundaries between skills and tools, or a central registry with proper version resolution are required.
The verdict is measured: Stitch Skills is a reference implementation, not a production-ready framework, and the standard is a proof of concept with real adoption. This is one developer's analysis, not a Google product announcement, and how widely the standard will be taken up is still unknown.
What to watch
- Will the Agent Skills standard add rules for retries, tracing and limits on which tools a skill may call?
- Will a registry or version-pinning approach emerge to replace plain Git sharing?
- How widely will other agent makers adopt the standard beyond the tools named so far?
Try it
Pick one task you repeat with an AI assistant and write it down as a mini skill: a name, a one-line purpose, the inputs it needs, the steps in order and what a good result looks like. Save it where you can paste it in.
Sources and links
- Reportingdev.to: the article we sourced
First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.