Dark Factories: Verification-First Agentic Engineering
summary
Claims that noninteractive runs show a 53% accuracy decline and that 91% of leaders report velocity gains but only 25% have data drive a verification-first design: an orchestrator decomposes requirements into a DAG planning pipeline, dispatches isolated sandboxed coding agents (containers/VMs) with Jujutsu worktrees, runs test executors plus plan adversary and code adversary gates, loops failed tasks through rework, and uses AWS (Bedrock, Lambda/Fargate, CodeBuild/CodePipeline, Step Functions, DynamoDB) for scale.
AI-written summary · about
chapters (6)
- 0:00What a dark factory means for software
- 4:03The convergent dark factory architecture
- 8:09DAG planning and rework loops
- 10:31A live factory from requirements to tasks
- 14:01Plan adversaries, code adversaries, and back pressure
- 16:57From writing code to composing the factory
related talks
46mPanel Discussion: The Agentic Software Factory That Enterprises Actually Need
51mMake Quality Executable: Verification Gates for Agentic Engineering at Scale
42mRestoring Trust in AI-Native Development
15mDevelop at Idea Velocity
20mBuilding Deterministic Infrastructure for Non Deterministic AI Agents
48mSecuring Agentic AI
29mBuilding Trust in AI Agents
30mArchitecting Integrity: Responsible Software Evolution in the AI Era