Skills That Self-Evolve With Every Use

The skills your agents use get better as your team works. Corrections become new versions, and each one ships only after it beats the last.

Book a demo See how skills evolve

Watch a Skill Evolve

Follow a sprint report skill from a suggestion in chat to its second version.

Product walkthrough, 90 seconds, following one skill through the six stages on three pages. Chat. Discover: someone asks for this week’s sprint status report. Nothing in the Skill Library fits, so recommendation mode suggests two skills from external repositories, jira-sprint-sync and sprint-status-report, and Details shows where jira-sprint-sync comes from and what has been checked. Install: the install guardrails fetch jira-sprint-sync, check its SKILL.md fields and read its repository’s scans, which pass, but our own scan finds it sends a Jira token to an unknown host, so it is blocked and nothing installs. sprint-status-report passes every check and installs privately at the version fetched. Use: the agent loads the skill for this request, pulls the sprint from Jira and writes the report, and a correction to put blockers first goes to memory. Skill Library. Manage: Your skills shows each skill’s source, version and update setting; an agent merges version 7 of Incident Report Drafting into an edited copy, keeping its edits; sharing the imported skill is blocked until it is reviewed, so it is submitted to the library. Evaluate: the evaluation gate checks format, security, triggering (18 of 20) and outcome (0.83 against 0.33 with no skill), and a reviewer who isn’t the author approves it as Sprint Status Report version 1. Memory. Evolve: a month later, memory shows 9 people made the same correction; an agent drafts version 2; it passes the gate and beats version 1, 0.86 against 0.83; a reviewer approves it, auto-update installs move to version 2, pinned installs stay on version 1, and version 1 stays available for rollback.

The Whole Skill Lifecycle

Six stages, from finding a skill to shipping a better version of it.

Stage 1 of 6

Discover

As people work, recommendation mode suggests skills from your Skill Library first, then from external skill repositories.

See recommendations

Discover Skills as You Work

Turn on recommendation mode, and chat suggests skills for the task you’re working on.

Searches Your Library First

If nothing there fits, it looks in external skill repositories.

You Choose What Installs

A suggestion installs only when you click Install, and it still has to pass the install guardrails.

The Exact Version You Fetched

We store the files you installed, so the skill changes only when you choose to update it.

Trust Signals Before You Install

Details shows what the repository, GitHub and your own organization know about a skill.

Install From Your Library or External Repositories

Library skills install in one click. Skills from external repositories must pass install guardrails first, and a failed check blocks the install.

Guardrails on Every Install

Each import is format-checked and scanned for secrets, risky scripts, network calls and hidden instructions before it can run.

Blocked When a Check Fails

A skill that fails any check never installs, and you see which check stopped it.

Private Until Reviewed

Imports stay private until they pass the evaluation gate and a named reviewer.

Checked Again on Update

Updating takes one click, and the update runs through the same checks before it installs.

Or Build a New One

When nothing fits, build a skill in chat with an agent, or write it yourself.

Build it with the skill builder agent

Describe the task in chat. The agent asks about your templates, writes the skill, and tests it with and without the skill before you save it.

---
name: internal-comms
description: A set of resources to help write internal communications …
---

## When to use this skill
- Status reports
- Leadership updates
- Company newsletters
- Incident reports

## How to use this skill
1. Identify the communication type
2. Load the matching guideline file …

Write it yourself

Write a SKILL.md file by hand in plain Markdown, with your own scripts and reference files.

New skills stay private until they pass the evaluation gate and a named reviewer.

Attach Skills Where You Work

Install a skill once, then attach it where you need it: to one chat, or to an agent or harness so every run can use it.

Attach to a Chat

Add a skill to the conversation you’re in, the way you attach a file. It stays with that chat.

Attach to an Agent or Harness

Add a skill to an agent’s setup once, and every run can use it.

Loads Only When Needed

An attached skill shows the agent only its name and description until a task needs it, so attaching many doesn’t slow anything down.

Harnesses Take Library Skills Only

Production harnesses attach library versions, which passed the gate and a named reviewer. Chats and agents take any skill you have.

Manage Sharing and Updates

A skill’s owner decides who can use it, and each person decides how their install updates.

Share With People or the Library

Sharing with specific people needs no review. Submitting to the library starts one.

Auto-Update or Pin a Version

Get each new version automatically, or pin one so nothing changes until you choose.

Customize Your Copy

Edit your own copy of a library skill. Updates never overwrite your changes.

Migrate With an Agent

When a new version ships, an agent merges it into your copy and shows you the diff to accept.

Evaluate Every New Version

A skill can make an agent worse by loading at the wrong time or giving poor instructions, so every version is tested before it ships.

Trigger Tests

About 20 prompts, half of them near-misses, each run several times with the skill alone and next to your other skills.

Outcome Tests

The same tasks run with and without the skill, graded by code checks, a model judge and a human sample. Cost and time are tracked too.

A Named Reviewer

Someone other than the author reads every file and change before release.

Self-Evolving Skills

A skill’s next version is drafted from what people corrected while using it, and it goes through the same evaluation gate as the first.

ChatPlatform team · New chat

Write this week’s sprint status report for the Platform team.

sprint-status-report
Loaded skillsprint-status-reportSKILL.md + 1 reference
jira · search_issuesProject PLAT, sprint 42
Sprint 42 · Platform

Done 18 · In progress 7 · Blocked 2. Progress by epic follows…

Put blockers first, and link each one to its ticket.

Saved to memory for sprint-status-report
1 / 5 · Use

Related Modules

Build
Agent Harness Builder
Assemble reusable enterprise skills and tools into specialized, multi-turn autonomous agents.
Governance
Evaluation Harness
Score agents and skills against a rubric before they ship, and catch regressions after.
Run & Deploy
Hosted MCP Servers
Deploy custom Model Context Protocol servers to expose enterprise APIs and databases to your tools.

Frequently Asked Questions

Common questions about Enterprise Skills & Tools.

What's the difference between a tool, a knowledge base and a skill?+
A tool lets an agent act on a system: send a Slack message, open a Jira issue or call an internal API. A knowledge base gives it your documents, such as Drive files, wikis, repos and policies. A skill is a set of instructions for a task, telling the agent which tools to use and which references to read, plus any scripts it needs.
What if the system we need has no API?+
There are two ways. With Model Context Protocol, you can build a custom tool for an internal API, a proprietary system, a database or a piece of business logic, and AI can write its definition and documentation. For a system with no API, browser automation works through its interface. It runs headless on Playwright and can navigate, fill in forms, scrape pages, export PDFs and run JavaScript.
How do we know a skill works before it goes live?+
Each new skill and each update goes through the evaluation gate: a format check, a security scan, trigger tests, and the same tasks run with and without the skill. A named reviewer who isn’t the author approves it, and the previous version is kept for rollback.
How do many skills load without slowing the agent down?+
An agent sees only each skill’s name and short description until a task needs it. Then it reads that skill’s instructions, and scripts run on their own so only their output comes back. So adding skills to the library barely affects each request.
Is it safe to use skills from external repositories?+
Every import passes install guardrails. It is fetched from its source, checked against the SKILL.md format, read against the security scans its repository publishes, and scanned by us for secrets, risky scripts, network calls and hidden instructions. If any check fails, it is not installed. We keep the exact version you imported, and it stays private to you until it passes the evaluation gate and a named reviewer.
What happens to our edits when a library skill updates?+
Your customized copy is never overwritten. When a new version ships, an agent merges it into your copy and shows you the changes, and you accept them or keep what you have. You can also pin any skill to one version.
Can a skill change itself without anyone approving it?+
No. An agent can draft a skill’s next version from memory, but it ships only if it passes the evaluation gate and beats the current one. Owners accept changes to their own skills, library skills need a named reviewer, and every earlier version stays available for rollback.