Keynote: The Self Driving Company
summary
Replit reports engineering output tripling over seven months while MTBF stays flat and MTTM drops; seven agentic loops automate coding, review, maintenance and incidents. A single engineer plus a loop produces 309 PRs to finish a stalled CSS migration; the review agent auto-approves low-risk PRs; nightly loops burn down flaky tests; a CTF-style harness finds dozens of vulnerabilities. Continuous improvement uses Bible Bench to validate changes and yields systems at 1/10 of the price of vendors.
AI-written summary · about
chapters (29)
- 0:00Start
- 0:05Speaker Intro - Luis Chavez, CTO of Replit
- 0:36Seven Months That Changed How Engineering Runs
- 1:06MTBF Stayed Flat and MTTM Actually Went Down
- 1:36What a Self-Driving Company Is - and Isn't
- 2:06Humans Set Goals; Agents Gather Context and Iterate
- 2:37Anatomy of the Agentic Loop and the Semantic Layer
- 3:08Loop One: Turbocharging Code Review, Not Just Coding
- 3:39Same Rules for the Coding Agent and the Reviewing Agent
- 4:10Letting the Agent Assess PR Risk and Auto-Approve Low-Risk Ones
- 4:41Loop Two: Maintenance - The CSS Migration That Had Failed Twice
- 5:11309 PRs Over Two Months, Written by One Engineer Plus a Loop
- 5:42Loop Three: The Networking Bug That Stumped Engineers for a Year
- 6:12Loop Four: A Nightly Run Burning Down Flaky Tests
- 6:44Loop Five: A Capture-the-Flag Harness for Penetration Testing
- 7:14Loop Six: Auto-Dispatching an Agent on Every Incident
- 7:44Loop Seven: The Agent That Improves the Agent
- 8:14Validating Improvements Against a Published Benchmark
- 8:45Opening It Up: Why Should Only Engineers Get Superpowers?
- 9:15The Harder Half - Non-Engineering Use Cases With No Test Suite
- 9:46In-Loop Validation: Tools, Guidelines and Grounding Evidence
- 10:16Never Jump to a Conclusion - Correlate the Metric Drop First
- 10:49Humans Around the Loop, Not Inside Every Loop
- 11:21Every Merged Correction Compounds Trust
- 11:51One Human Learning vs the Whole Company Learning Instantly
- 12:23Token Maxing vs Outcome Maxing - and Why They Chose Outcomes
- 12:53Minimising Time to Outcome, Not Maximising Spend
- 13:24Testing Was a Phase; Now It's a Property of Every Loop
- 13:55Loops Designed by People Who Think in Quality Produce Better Results
related talks
58mKeynote: Seven Habits of a (Mostly) Successful Team
49mKeynote: Operations and AI with Jen Krieger
49mKeynote
55mKeynote: Exoskeletons, not Autopilots
56mKeynote: CTRL+SHIFT+(BUILD) PAUSE
48mWe're Good At Writing Software [Keynote]
31mKeynote: How Software Testing Can Increase Agent Autonomy
54mKeynote: AI-Powered App Development