When You Run Out of Requirements: What Happens When You Go All-In on Agents
summary
An engineering team goes all-in on agents using Claude Code and Cursor plus custom HIPAA-tuned skills and an agent-first architecture; autonomous builds run plan‑act‑observe‑verify loops inside a four-part harness (context, tools, guardrails, memory). Feature cycle time drops 99% from days to minutes, PR throughput rises 4–5x and stays stable for a quarter, but the bottleneck moves upstream to product and UX because judgment and defining “done” remain human tasks.
AI-written summary · about
chapters (48)
- 0:00When You Run Out of Requirements: Going All-In on Agents
- 2:06Speaker Intro - Rashi Agrawal, Head of AI, Hinge Health
- 2:36The Backlog Disappeared - and That Was the Confusing Part
- 3:07Three Sprints of Backlog, Then Claude Code, Cursor and Custom Agents
- 3:37Four Months Later the Problem Didn't Go Away - It Moved
- 4:08The AI SDLC in 2026: AI Embedded in Every Phase
- 4:41231 Companies Across 16 Categories, Built in Three Years
- 5:4197% of Teams Using AI; 76% Say More Engineers, Not Fewer
- 6:44The Real Production Stack and Healthcare-Specific Skills
- 7:14Agent-First Instead of Adding AI to Existing Workflows
- 7:45Opening the Hood: How the Autonomy Actually Works
- 8:15The Inner Loop vs the Outer Loop
- 9:16Handing the Entire Outer Loop to Agents
- 9:47Plan, Act, Observe, Verify, Repeat - The Engine
- 10:18PR Cycle Time Down About 99%
- 10:50The Model Is Not the Hard Part - the Harness Is
- 11:20Four Parts of a Harness: Context, Tools, Guardrails, Memory
- 12:22The Model Is a Commodity; The Harness Is Your Moat
- 12:52Vibe Coding vs Agentic Engineering - and Why It Matters in Healthcare
- 13:54One Engineer Supervising a Fleet, and 4-5x PR Throughput
- 14:54Autonomy Is Earned, Not Granted: Three Tiers
- 15:56Tiering by Risk - More Rope for Low-Risk Internal Changes
- 16:26The Twist: The Constraint Moved Upstream, Past Design
- 16:58Writing the Spec Was Never the Hard Part
- 17:28Judgement Is the One Thing That Didn't Get Faster
- 18:0073% Hit Problems From Accepted AI Code Nobody Understood
- 18:32Speed and Quality Are Two Different Dials
- 19:03Someone Still Has to Define What Good Means
- 19:35AI Can Generate Thousands of Tests - It Can't Tell You Which Matter
- 20:37Three Threads: What to Build, What's Safe, What Done Means
- 21:09The New Bottleneck Is Judgement Itself
- 21:39Superpower One: Context Architecture
- 22:10The Test Coverage Example: 47 Targeted Edge Cases From Real Context
- 23:43Superpower Two: Designing the Verification Machine
- 24:16A Four-Tier Model From Auto-Approve to Dual Review Plus Audit
- 25:20Superpower Three: The Judgement Call Itself
- 25:5271% Won't Merge AI Code Without a Human Check
- 26:54Superpower Four: Lived Human Experience
- 27:26What the Agents Missed: Device Switching, Time Zones, Emotional State
- 28:29These Skills Aren't Dying - AI Freed You From the Other Ones
- 29:30This Is a Promotion If You Take It
- 30:00The Six-Month Path: One Tool Daily, Then the Team, Then Product Meetings
- 31:00The QA Engineer of 2027 and What Their Day Looks Like
- 32:3287% Say AI Adoption Future-Proofs Their Career
- 35:10Q&A: Safe Sandboxes and Discouraging Shadow AI
- 37:16Q&A: Where the Greatest Growth in QE Will Be
- 39:19Q&A: Compliance When Static Specs Give Way to Dynamic Behaviour
- 42:26Q&A: Do You Have to Progress Through Every Autonomy Tier?
related talks
40mHow To Tame Your Agents?
15mDevelop at Idea Velocity
45mAI agents in practice - beyond the prototype
30mArchitecting Integrity: Responsible Software Evolution in the AI Era
46mHarness Engineering: How to Build Software When Humans Steer, Agents Execute
27mAgentic Engineering: Working With AI, Not Just Using It
30mAgents Building Agents
21mThe Agent Native Company