Agentic workflows that compound
summary
Argues that agentic workflows compound by combining three concrete practices: a two-layer persistent memory (repo status.md snapshot plus a self‑hosted Munin (Raspberry Pi) with MCP and vector search), automated session close that writes curated notes and nightly syntheses, and routine adversarial debate with Codex (multi‑model, fresh contexts) that yields measurable critique. Shares production data showing Codex finds about 4x as many issues as the primary reviewer, describes cost and token‑management tactics, and shows how metrics drive incremental tooling improvement.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 2:37Speed gap in agentic dev
- 4:46Three pillars for compounding
- 6:55Persistent context with Munin
- 10:06Adversarial review for errors
- 16:48Defining workflow quality
- 21:49Summary
- 22:10Q&A
related talks
44mBuilding Agentic Workflows in Cursor IDE
32mFigma deep dive: Agentic workflows and the MCP
27mAgentic Coding with Discipline & Skill: Practical Techniques and Tips
35mAgentic Workflows: Stories from the Engine Room
16mAgents vs Workflows: Why Not Both?
23mAgentic Elixir
17mSoftware Development Agents: What Works and What Doesn't
29mFrom Copy-Paste to Composition: Building Agents Like Real Software