AI Agents That Don’t Suck: How to Build Ones That Actually Drive Business Value
summary
Quartzip builds a voice onboarding agent using WebRTC, Cloudflare, Kubernetes and a pipeline that moved from Google Gemini multimodal to transcription + Anthropic Claude via PipeCat and Daily, cutting calls from 45 minutes to 5 minutes and cold-start latency from 30s to 5s with caching. The talk shows concrete failures (80%+ AI project failure rates, many pilots with zero ROI), and prescribes narrow workflow scope, validation gates/structured outputs, eval harnesses in CI/CD, and cost/scale tradeoffs to avoid token and memory blowouts.
AI-written summary · about
chapters (14)
- 0:00Building useful AI agents
- 1:32AI technology is solved
- 3:57User psychology and business value
- 9:59Why AI agents fail
- 12:45Guardrails and system design
- 22:36Users want holes not drills
- 29:58Engineering better agents
- 36:42Reducing agent complexity
- 41:56Context engineering and caching
- 47:17Evaluating agent performance
- 50:11Security risks in agents
- 51:14Iterative agent improvement
- 52:18AI agent frameworks
- 53:20Next steps for agents
related talks
43mTesting Agentic AI Applications: Beyond Traditional QA
30mAI Agents Unchained
49mHow to be an AI First company
18mFrom POC to Profit: Deploying Customer-Facing AI Agents That Actually Work
36mAI Agents in Automation | 3 Real Production Cases with ROI
20m10 Things I Learned Building Voice Agents in 2025
26mFour Steps to Embed AI Agents into Workflow
17mBuilding Effective Voice Agents