Directing a Swarm of Agents for Fun and Profit
summary
Directs how teams of AI agents actually build software: uses GasTown/Roof Flow for multi-agent specialization (testing, architecting, researching), bakes conversational history into a block comment at top of each file to preserve intent, prefers given-when-then BDD for fewer fake tests, and runs fractional-factorial design-of-experiments to evaluate model/language/tool permutations. Demos a live repo generator called Retort benchmark tool that spun up a stack this morning, references BrazilBench, 150,000 lines of Python, a working mobile client and an iMessage Insta integration, and notes a 50,000‑star paperclip repo showing rapid zero‑human org emergence.
AI-written summary · about
chapters (8)
- 0:00Introduction and Framing
- 1:14Why Play Matters for Profit
- 3:03NoOps to NoDev Parallels
- 4:19Agentic Coding Techniques
- 7:03Multi-Agent Systems and Projects
- 10:50Making Systems Conscious
- 18:24No-Manager Companies and Paperclip
- 19:38Live Demo: Retort Benchmark Tool
related talks
24mAI Agent Swarms Are Amazing
44mThe Swarm Next Door: Running Parallel AI Agents on a Real Android Codeba…
39mAgents for the Sake of Happiness
38mBuilding AI Agents and Assistants for DevOps Workflows
40mHow To Tame Your Agents?
45mAI agents in practice - beyond the prototype
25mBeyond the IDE: Toward Multi-Agent Orchestration
18mCollaborative AI Engineering: One Dev, Two Dozen Agents, Zero Alignment