Intelligence Snacks podcast

Intelligence Snacks 79 - Cloud Agents

0:00
1:02:59
Spola tillbaka 15 sekunder
Spola framåt 15 sekunder

Can AI agents work for you without having a permanent computer? Pete and Andy explore cloud agents, temporary computing environments and why an agent's identity, memory and tools can persist even when its computing resources don't. Pete explains his work on Wingman Lite and agent swarms through Ambulando and NStack, and how shared infrastructure could make capable agents cheaper and easier to use.

The conversation also turns to how people actually work with agents. Pete's Kanban boards are becoming records of work rather than instructions for humans, while Andy is asking Lara to supervise tasks and orchestrate his Snack Studio pipeline. They discuss building apps before knowing what you want, why some software becomes unnecessary after you've built it, and when you still need to work closely alongside an agent.

From there, Pete and Andy examine the limits of human supervision, information security between personal and shared agents, and what happens when AI makes execution and compliance dramatically cheaper. Could smaller companies take on projects once reserved for much larger organisations? And how much of what's possible are we missing because we're still thinking in terms of yesterday's constraints?

Key Moments:

  • [01:19] Welcome to Episode 79 and why agents are changing Kanban boards.

  • [03:22] Using Lara to triage another agent's tasks.

  • [04:32] Pete's visual task trees and shaping work without writing every task.

  • [06:54] Does every workflow need its own app?

  • [08:46] Activity dashboards, posting more and changing habits.

  • [09:40] Building a BD app before knowing what it should look like.

  • [11:43] Why it's easier to say what's wrong than specify what's right.

  • [14:57] Snack Studio and discovering you don't need the app you built.

  • [17:23] Why full agent workspaces still have a place.

  • [19:16] Managing agents versus working alongside them.

  • [20:51] Pete's children and AI-powered research wikis.

  • [23:44] Wingman Lite and easier access to local agents.

  • [26:18] Cloud agent swarms that spin up only when needed.

  • [28:38] Shared infrastructure and the economics of agent computing.

  • [30:11] How agent identity persists without a permanent computer.

  • [32:37] Persistent computers versus isolated cloud agents.

  • [35:16] Personal and shared agents, identity and data access.

  • [39:09] Event-triggered agents that produce reports and shut down.

  • [41:38] Why permissions need to make sense to humans.

  • [44:09] Can human supervision keep up with agent swarms?

  • [45:38] Why AI slop is a problem of care and attention.

  • [47:47] Falling legal costs and new opportunities for smaller businesses.

  • [50:12] Morphic fields, intuition and how things take shape.

  • [54:50] Why changing constraints could reshape whole businesses.

  • [56:52] "Why not us?" and unlearning limits on ambition.

  • [59:33] How AI could reduce the advantage of big organisations.

  • [01:02:08] Being early without running out of money.

Projects and ideas mentioned: Ambulando, NStack, Wingman, Wingman Lite, Autopilot, Snack Studio, Lara, Rick, Kindling, Excalidraw, Manus, Muse, cloud agents, agent swarms and morphic resonance.

Find out more: ambulando.io, intelligencesnacks.com

Closing thought: "Be early, be solvent." — Pete

Fler avsnitt från "Intelligence Snacks"