Agentic engineering with Cosmos
Coding is only about 30% of an engineer's job. The rest is the "assembly line": review, testing, tickets, on-call, and design. According to Amdahl's Law, if you only optimize the coding step—even by 100x—your total throughput speedup tops out at around 43%. To get to 3x, you have to automate the entire line. In this talk, lead engineer Akshay Utture (PhD UCLA, ex-Uber) explains how Augment uses Cosmos to automate the high-latency stages of the SDLC. He walks through a real-world case study: how a 1,000-line PR that used to take seven hours of reviewer time now merges in 45 minutes. Key Technical Deep Dives: + The Code Review Fleet: How four distinct agents (Deep Code Review, PR Author, Risk Analyzer, and Pair Reviewer) handle correctness, auto-fixes, and risk assessment before a human reads a line. + Managing Runtime Breaks: A look at how the Deep Code Review agent caught a high-severity back-compat issue involving "include" resolution—a break that standard linters and basic LLM prompts miss. + Auto-Approval Thresholds: Why we started with a conservative 10-15% auto-approval rate for low-risk changes, and how the Risk Analyzer determines which PRs don't need a human in the loop. + Incident Management at 3x Velocity: How the Incident Investigator agent reduced median RCA time from 30 minutes to 6, handling 81% of incidents without manual triage. + The "Judgment" Shift: What engineering work looks like when agents handle the mechanical work (reading diffs, log running, commit crafting) and humans focus on architectural tradeoffs and expert tuning. The Numbers: + Weekly Code Output: 1x to 3x + PR Merge Time: ~700 mins down to 250 mins + Bug Rate: 0.097 to 0.006 commits per output unit + On-Call Impact: 44% more PRs merged during on-call weeks About the Speaker: Akshay Utture leads the engineering efforts behind the code review agents at Augment. His PhD at UCLA focused on making static analysis practical for code review, and he previously spent time on developer tooling at Uber. #SoftwareEngineering #SDLC #AI #CodeReview #Cosmos #DevOps #EngineeringManagement #Root #EngineeringManager #EngineeringProductivity