Agent Omelets & Other Forbidden Techniques
summary
Shows concrete engineering moves: you can roll back and fork live agent state to resurrect earlier behavior, run hydro runs that fork the same context across Anthropic/OpenAI/other providers and have an agent pick the best output, edit an agent’s memory (and erase the trace), hot‑swap model providers mid‑conversation, and attach an omelet reference to a posted brief that lets you resuscitate the exact agent state later. Mentions an "internet for agents" (documents, rosters, handoff briefs) and a packaged supply‑chain product (ChainGuard libraries) for safer agent code execution.
AI-written summary · about
chapters (14)
- 0:00Introduction and setup
- 1:11Forbidden techniques concept
- 3:38Rollback and forking agents
- 6:49Memory editing and hydro runs
- 10:45Introducing the informational layer
- 13:56Ignition and torch process
- 14:18Agenttology and agent profiles
- 16:43Shadow cloning and resuscitation
- 18:08Agent integrity and isolation
- 19:52Chain Guard libraries overview
- 21:18Sub agents and agent conversations
- 22:22Pokey and tool design
- 23:24Embrace agent weirdness
- 23:45Vibe Olympics event
related talks
40mHow To Tame Your Agents?
01h 26mAI Agents of Change: Creating, Reflecting, and Monetizing
49mAI Agents: Collaboration in Real-Life Applications
26mBeyond Hype: Leading the Charge in Responsible AgentOps for Business Transformation
30mAI Agents Unchained
01h 01mAI Agents In-Depth – Function Calling, MCP and Tool Use Under the Hood
29mAgent-Ready Orchestration: Engineering the Execution Layer.
33mAgent Experience