Agent and skill distribution platform
Turned copy-pasted AI agent files into versioned packages from a single registry, with a cross-platform installer and a staleness report.
- Result
- 18 repos
on versioned agent packages, from 7 packages - Role
- Lead engineer; designed the registry, generator, installer and rollout
- Domain
- National telecom carrier
- Period
- May 2026 – present
Problem
Test-automation teams were adopting AI agents by copying agent and skill files between repositories. Nobody could tell which copy was current, fixes didn't propagate, and there was no review gate.
Approach
Treat agents like software packages: one registry as the source of truth, everything else generated, and a rollout that works on locked-down Windows laptops.
What I built
- A
packages.registry.jsonsource of truth. A generator emits every manifest, doc and README table, and validate and smoke scripts gate every change through a pre-commit hook (CI runners weren't available). - Seven packages: web test generation, self-healing, Selenium-to-Playwright migration, PR review, API test generation, version management, and a shared default.
- A one-run installer for Windows and macOS that bootstraps the CLI toolchain and lets engineers pick packages.
- A rollout plan (one bootstrap PR per repo, then self-update) and a staleness report across consumer repos.
Outcome
- About 18 consumer repositories on versioned packages instead of copies.
- Rollout hardened against real enterprise constraints: Windows path limits, Mark-of-the-Web, and policies that disable agent hooks.