Agentic Engineering: From Writing Code to Reviewing It
summary
Describes an operating model where AI agents accept a ticket and deliver a plan, tests and a reviewable PR rather than a keystroke; a demo shows a PR merged in 11 minutes that no human wrote, another shows an autonomous investigation confirming 600 failures and delivering a fix in about nine minutes, and a security product farms work to agent swarms with lead and worker agents. Emphasizes the L0–L5 autonomy ladder, local vs cloud agents, and the golden test: can success be verified without human judgment.
AI-written summary · about
chapters (40)
- 0:00Agentic Engineering: From Writing Code to Reviewing It
- 1:37Speaker Intro - Sumit Shatwara, Deployed Engineer at Cognition
- 2:08The PR That Was Merged in 11 Minutes and Written by Nobody
- 3:09Putting Coding Agents Inside the Most Regulated Codebases
- 4:42The Autonomy Ladder: L0 to L5 and the Line in the Middle
- 5:46Left of the Line Augments an Individual; Right Multiplies a Team
- 6:46L3: Agents With Their Own Machine and Real Feedback Loops
- 7:49Every Rung Removes a Human Checkpoint - Raising the Value of Review
- 8:21The Unit of Software Work Shifts From Keystroke to Ticket
- 9:24Three Things Become Scarce: Intent, Judgement, Quality Review
- 9:54Local Agents vs Cloud Agents - Two Mental Models
- 11:27The Economic Asymmetry: Hours of Writing Into Minutes of Reviewing
- 11:59"You Are Not the Driver Anymore, You Are the Reviewer"
- 12:31Why Engineers Fail at Delegation - Micromanaging the Agent
- 13:34The Golden Question: Can Success Be Verified Without Human Judgement?
- 14:34If Yes, It's a Strong Agent Candidate - If No, It Needs Your Judgement
- 16:06Your Test Suite Is the Speed Limit on Your AI Adoption Curve
- 17:07Why QE Has a Much Larger Role in the Agentic World
- 17:37Demo 1: Delegating a Well-Scoped Ticket With Success Criteria
- 19:08Watching Devin Plan, Edit, Build and Open a PR in an Isolated VM
- 20:38End-to-End Visual Validation and a Full Reviewable Package
- 22:10Removing Yourself From the Trigger Entirely
- 22:40Three Field Patterns: Vulnerability Remediation, Bug Investigation, Slack Monitoring
- 24:13Moving the Human From the Trigger to the Final Gate
- 24:44Demo 2: A 3 AM Production Alert in a Banking App
- 26:14Sentry to Slack to Autonomous Investigation and Root Cause
- 27:46600 Failures Confirmed, Fix Written, Tests Run - in Nine Minutes
- 30:21Scaling Up: Agent Swarms With a Lead and Worker Agents
- 31:54We Didn't Automate Engineering Judgement - We Maximised It
- 32:58Demo 3: The Security Swarm Fanning Out Across a Codebase
- 33:59Aggregation, Deduplication and False-Positive Validation
- 35:38The Full SDLC Map: DeepWiki, Ask Devin, Cloud, Review, Automations
- 37:11Watching the CI and Re-Running Until It Passes
- 38:15The Question Every Tester Is Asking: What If All Three Gates Pass?
- 38:46Demo 4: The 5:55 PM Prompt That Says "Just Make CI Green"
- 39:48The Agent Flags That It Was Told to Overwrite the Source of Truth
- 40:18Devin Review Catching Three Bugs Including Non-Latin Titles
- 41:53Typical Use Cases: Modernisation, Tech Debt, Coverage, Vulnerabilities
- 42:54Delegate One Verifiable Task - Then a Second
- 43:24Q&A: How Long a Full SDLC Cycle Takes; Mobile and Game Use Cases
related talks
27mAgentic Engineering: Working With AI, Not Just Using It
35mThe Agentic AI Engineer
29mPairing with the machine: How AI Agents are rewriting the rules of software development
25mFrom AI Chaos to Production Ready: Using AI Agents to Improve Your Code Quality
01h 07mNew Methods of Working with AI.
26mAI Agents as Your Co-Developer
01h 00mKenote: AI Didn’t Replace You. It Promoted You!
30mArchitecting Integrity: Responsible Software Evolution in the AI Era