Stenobird

Episode

AI Barista Orders 120 Eggs With No Stove, Matt Pocock Skills, Build for the Next Model, The Gervais Principle

Podcast
You've Been a Bad Agent
Published
May 9, 2026
Duration seconds
3018
Processing state
processed
Canonical source
https://share.transistor.fm/s/f365174e
Audio
https://media.transistor.fm/f365174e/d6f2eac0.mp3
JSON
/v1/public/podcasts/you-ve-been-a-bad-agent-7334728/episodes/ai-barista-orders-120-eggs-with-no-stove-matt-pocock-skills-build-for-the-next-model-the-gervais-principle
Markdown
/podcast/you-ve-been-a-bad-agent-7334728/ai-barista-orders-120-eggs-with-no-stove-matt-pocock-skills-build-for-the-next-model-the-gervais-principle.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/you-ve-been-a-bad-agent-7334728/episodes/ai-barista-orders-120-eggs-with-no-stove-matt-pocock-skills-build-for-the-next-model-the-gervais-principle/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/you-ve-been-a-bad-agent-7334728/ai-barista-orders-120-eggs-with-no-stove-matt-pocock-skills-build-for-the-next-model-the-gervais-principle.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

An exploration of the friction between rapid AI capability gains and the messy reality of real-world implementation. The hosts discuss everything from autonomous agents making disastrous supply chain errors to the potential for 'skills' to act as a new prompt-injection attack vector.

Topics

  • AI Agents
  • LLM Prompting
  • Automation
  • Software Engineering
  • Autonomous Systems
  • Cloudflare
  • Anthropic
  • Prompt Injection

Highlights

  • Failure mode: Autonomous agents can cause physical-world chaos, such as ordering massive quantities of unusable supplies due to a lack of environmental context
  • Main idea: The 'capability overhang' remains significant; even as models scale, the gap between raw intelligence and reliable, low-stakes decision-making is wide
  • Practical takeaway: Using structured 'skills' (markdown paired with deterministic scripts) can significantly improve model performance and reliability
  • Security risk: New prompting techniques like 'skills' may inadvertently create new prompt-poisoning attack vectors by being treated with system-prompt authority
  • Future outlook: The next era of work may shift from performing tasks to becoming 'automation engineers' who manage agentic workflows

Chapters

  1. 1:00 Reflections on a Year in SF: A look back at the past year of the podcast, moving from startups to Cloudflare and the unique, time-distorting atmosphere of San Francisco.
  2. 16:00 The Slow Pace of AI Progress: Discussing the tension between massive ARR growth at companies like Anthropic and the persistent difficulty in reliable prompting.
  3. 24:00 The AI Barista Hall of Shame: A deep dive into the errors of Mona, an AI agent at a Stockholm cafe, including the infamous 120-egg order.
  4. 27:00 The Future of Automation Engineering: Debating whether employees will eventually be tasked solely with automating their own roles and the implications for the workforce.
  5. 39:00 iOS Live Activities and UI Surfaces: Exploring how mobile OS features like Live Activities can serve as high-priority surfaces for agentic notifications.
  6. 42:00 Building for the Next Model: The frustration and strategy of developing tech that works on current models but might break with the next iteration.
  7. 46:00 Prompt Poisoning and Scripted Skills: Analyzing how Matt Pocock's 'skills' approach uses markdown and scripts to change model behavior and the potential security risks involved.