AI Agents Unchained
summary
A runbook of concrete wins and failures: a 2018 Chrome‑extension system (patented as the AI Communications Agent) actually automates phone lookups and payment options for a telco; an OpenClaw deployment deletes real emails when auto compaction drops a “don’t act” instruction and a user must sprint to unplug it; a multi‑agent SDR stack uses an orchestrator pattern with heartbeats that improves lead flow but raises who-to-ask and failure‑point questions; overall claim: harnesses, not models, often make or break systems (uploader cites 95% fail attempts).
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:36OpenClaw email deletion
- 5:36AI agent harness components
- 8:28Multi-agent SDR case study
- 11:20Orchestrators vs supervisors
- 13:07Autonomy versus safety
- 14:30Claude Code harness failures
- 17:00Devin's harness improvements
- 19:50Risks of over-engineering
- 22:46Evaluating harness impact
- 24:57Earned competence in design
- 25:40Q&A
related talks
24mAI Agent Swarms Are Amazing
46mAI Agents Are Becoming Operating Systems: Lessons From OpenClaw
34mAI Agents: Transforming Autonomous Decision-Making
49mAI Agents: Collaboration in Real-Life Applications
56mAI Agents That Don’t Suck: How to Build Ones That Actually Drive Business Value
44mPannel Discussion | Agentic Automation: Are AI Agents The Future of Digital Work
33mAgent Experience
26mAI Agents: How Do They Work and How to Build Them