Stenobird

Episode

Sub Agent or Super Agent, MCP UI Over Lunch, Bitter Lesson Learnings, Locking in on Prompts and Trying to Live in the Future

Podcast
You've Been a Bad Agent
Published
Aug 6, 2025
Duration seconds
3466
Processing state
processed
Canonical source
https://www.youtube.com/watch?v=8moSHUSSwKI
Audio
https://media.transistor.fm/7753e1ba/7a100409.mp3
JSON
/v1/public/podcasts/you-ve-been-a-bad-agent-7334728/episodes/sub-agent-or-super-agent-mcp-ui-over-lunch-bitter-lesson-learnings-locking-in-on-prompts-and-trying-to-live-in-the-future
Markdown
/podcast/you-ve-been-a-bad-agent-7334728/sub-agent-or-super-agent-mcp-ui-over-lunch-bitter-lesson-learnings-locking-in-on-prompts-and-trying-to-live-in-the-future.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/you-ve-been-a-bad-agent-7334728/episodes/sub-agent-or-super-agent-mcp-ui-over-lunch-bitter-lesson-learnings-locking-in-on-prompts-and-trying-to-live-in-the-future/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/you-ve-been-a-bad-agent-7334728/sub-agent-or-super-agent-mcp-ui-over-lunch-bitter-lesson-learnings-locking-in-on-prompts-and-trying-to-live-in-the-future.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

The future of AI development lies in managing context through sub-agents and standardized protocols like MCP. The discussion explores the tension between powerful models and the practical difficulties of prompt engineering and tool integration.

Topics

  • AI Agents
  • Model Context Protocol
  • Prompt Engineering
  • Software Engineering
  • Sub-agents
  • MCP Servers
  • AI Tooling
  • Context Windows

Highlights

  • Main idea: Sub-agents are essential for managing context windows and preventing main agent pollution
  • Practical takeaway: Developing better UI for prompt engineering is more valuable than creating new SaaS-based prompt management tools
  • Failure mode: Relying on custom prompting as a competitive moat is risky in an era of rapidly advancing foundational models
  • Technical insight: MCP (Model Context Protocol) servers provide a standardized way to extend agent capabilities via specialized tools
  • Practical takeaway: Effective AI-assisted coding requires robust feedback loops and live evaluations to overcome the volatility of model performance

Chapters

  1. 1:00 Outdoor Adventures: A brief look at recent fishing and water sports experiences in Dorset.
  2. 10:00 Context Management and Sub-Agents: Analyzing how sub-agents can be used to manage context windows and specialized rules without overwhelming the primary model.
  3. 14:00 The Role of MCP in Agent Architecture: Discussing the use of MCP to provide sub-agents and the potential limitations of centralized agent control.
  4. 18:00 Tooling and Permissions: Examining the complexities of managing actions, permissions, and disallowed tools within agentic workflows.
  5. 23:00 The Volatility of AI Coding: Reflecting on the inconsistent performance of AI coding tools and the difficulty of maintaining control over prompts.
  6. 27:00 Prompt Engineering and UI: Critiquing current prompt writing workflows and the need for better structured, visual interfaces for prompt development.
  7. 31:00 Future-Proofing AI Skills: Discussing the importance of mastering tools and form factors that will remain relevant in the next five years.
  8. 40:00 The Future of Agentic Workflows: Speculating on the shift toward highly automated, agent-driven development and the potential for new technological frontiers.