Case study: a dev agency shipping client work with AI agents, safely

By Shamaz Khan · 18 Nov 2026 · 6 min read

Illustrative scenario. The team and numbers below are a composite example, not a real customer. The problem and the fix are real.

TL;DR

The situation

A six-person agency maintains twelve client apps. Since moving to Claude Code and Cursor, they ship maintenance work twice as fast. They also had three incidents in two months where a "small" change broke something a client relied on: a report export, a booking confirmation email, a login redirect.

Why agencies feel this more

What they standardised

Every client repo gets the same baseline

npx stet-cli init --ci

That sets up the rules files and writes the pull-request check. Then each account lead locks the three behaviours the client cares most about:

stet lock add "Booking confirmation email is sent" --cmd "npm test -- tests/booking-email"
stet lock add "Monthly report export columns" --snapshot "node scripts/print-report-columns.js"
stet lock add "Login redirects to the dashboard" --cmd "npm test -- tests/auth/redirect"

Every task gets a scope

stet scope set "ACME-212 Add VAT field to invoice form" --allow "src/invoices/**,tests/invoices/**"

The ticket ID flows into commit messages and the PR description automatically.

Every PR shows its evidence

stet pr --what "…" writes the description from the diff, the scope result and the lock results, so the client's reviewer sees which files changed, that everything stayed in scope and which behaviours were checked.

The result

Out-of-scope edits get caught at commit time instead of by clients. Just as useful: the agency now shows the scope and lock results in client updates, and "every change is checked against the behaviours you care about" has become part of their pitch.

FAQ

Is it safe for agencies to use AI coding agents on client projects?

Yes, with guardrails: a scope for every task, locks on the behaviours each client depends on, and a check on every pull request so regressions are caught before the client sees them.

How can an agency prove AI-written changes are safe?

Put the evidence in the pull request: which files changed, whether they stayed in scope, and which locked behaviours passed. stet pr generates that from facts.

What should an agency lock in a client app?

The two or three behaviours the client would notice first: emails they send, exports they use, sign-in and anything involving payments.

Put guardrails on your AI coding agent

Stet is a free, open-source CLI and MCP server for Claude Code, Cursor, Codex and Copilot. It runs locally: no account, your code never leaves your machine.

npx stet-cli init · Install · Setup guide · GitHub

Related