The AI-Native Development Loop: A Ticket Comes In, a Verified Change Goes Out

From BrowserStack Meetup Cape Town (BBD, Century City, 3 Sep 2026) — Introduction to AI-Native Development Workflows.

A human-gated engineering loop: Linear ticket via MCP → reproduce → AI-assisted fix (model-agnostic: Claude, Cursor, Grok) → lint / FE+BE / Playwright → BrowserStack real devices via Playwright SDK + Local tunnel → human check → PR → report back to Linear. Not full automation. Three human gates (pick it up, reproduce, ship). Cheap local proof before cloud devices. MCP is for Linear only.

Event: https://luma.com/browse-6tdk

Transcript

1. BROWSERSTACK MEETUP · CAPE TOWN AI-NATIVE DEVELOPMENT The loop. A ticket comes in. A verified change goes out. Human in control · not full automation Abed Matini BBD, Century City · 3 Sep 2026
2. THE IDEA Work should move in a circle, not a pile. 01 02 03 Arrive Address Report Task lands in Linear. We read it through MCP. Reproduce. Fix. Prove it — locally, then on devices. PR goes up. Linear gets the story of what happened. AI-native loop 2 / 13
3. THE CIRCLE Six steps. Same order every time. 1 2 3 4 5 6 Intake Reproduce Fix Prove Devices Close Linear · MCP Does it exist? Any model, you drive Lint · tests · E2E BrowserStack PR · report Local proof comes before BrowserStack. Then a human checks the ticket is actually done. AI-native loop 3 / 13
4. 01 · INTAKE Read the ticket. Don’t code yet. MCP to Linear Pull the issue, description, comments, attachments. Is this a code task? Questions, design, and product asks skip the loop. Enough to start? Missing repro, env, or acceptance → send it back. AI-native loop 4 / 13
5. 02 · REPRODUCE Does the problem exist? If we cannot see it, we do not invent a fix. Bug Feature / gap Hit the path. Same browser, same data, same flag. Screenshot or trace. Confirm the current behaviour and write the acceptance in one line. Human gate A person confirms: this is real, this is in scope, this is worth a fix. AI-native loop 5 / 13
6. 03 · FIX Any model. You still drive. Model-agnostic Ticket-sized change Not auto-merge Claude, Cursor, Grok — same loop. Only what the issue asked. Nothing heroic. A person ships. Always. AI-native loop 6 / 13
7. 04 · PROVE LOCALLY Green here, or we do not go to devices. Lint FE + BE Playwright Style and obvious mistakes. Unit and integration suites. E2E on the happy path. Flaky E2E is a loop-killer. Treat flakes as bugs, not as retry-until-green. AI-native loop 7 / 13
8. 05 · DEVICES Real browsers. Real phones. Evidence. SDK + Local tunnel Same Playwright suite. Tunnel to BrowserStack — not MCP. A small matrix A few viewports and devices. Not every phone on earth. Keep the recording Video + result go back on the Linear issue. AI-native loop 8 / 13
9. 06 · CLOSE THE LOOP Look once. PR. Tell Linear. 1 2 3 Human look Raise the PR Report back Does this actually address the ticket? Not just green. Diff, test evidence, BrowserStack links. Comment + status on Linear. PR linked. What changed. AI-native loop 9 / 13
10. HUMAN IN CONTROL Three gates. Everything else can move. Gate 1 Gate 2 Gate 3 Pick it up Reproduce Ship Is this a code task we should even start? Did we actually see the problem? Does the PR match the ticket? AI-native loop 10 / 13
11. WATCH-OUTS The loop is right. These will break it. Not every Linear ticket is a code task. Green tests are not the same as the ticket being done. Flaky Playwright will lie to the loop. BrowserStack after local green — never instead of it. Linear MCP writes can miss. Check the comment actually landed. Never auto-merge. Not in this version. AI-native loop 11 / 13
12. LIVE Demo. One ticket. All the way around.
13. THANK YOU Same circle. Tighter gates. More automation later. Not yet. Abed Matini linkedin.com/in/matini BrowserStack Meetup Cape Town · BBD Century City