I Let an AI Agent Buy My Pants
Episode ten: Jason Patel splits AI builds into planned projects and "copy and customize", where the tokens disappear fast. From there: Linear versus markdown files, giving each sub-agent its own Linux user, why agents don't need email, and whether engineers will make good team leads of agents. Matt and Jason compare how much access to give Claude, ask who clicks past an AI summary, and talk LLM tutors, messy government data, and hallucinations versus mistakes. Plus Matt's German learning app and the AI he let buy him pants.
Episode ten. Jason Patel splits AI builds into planned projects and “copy and customize”, and he and Matt compare how they actually work with AI coding agents.
What we talked about:
- 0:00 Cold open: I just give it full access
- 0:31 Planned builds vs. copy-and-customize
- 2:28 Months of planning, then constant replanning
- 4:57 Markdown files, Linear, and worktrees
- 6:43 Giving every sub-agent its own Linux user
- 10:05 Email for agents?
- 11:44 Where the token burn comes from
- 13:50 UI work burns the most tokens
- 16:47 Testing a Mac app in the cloud
- 18:21 A bigger terminal font fixed my headaches
- 19:47 Agents aren’t employees
- 21:23 Will engineers be good team leads of agents?
- 23:34 What happens to the ticket closers?
- 25:40 Dev and QA become one again
- 28:44 The adapter pattern, but for Claude
- 30:48 How much access do you give Claude?
- 35:33 Go-to-market with AI
- 38:37 Does anyone click past the AI summary?
- 43:33 I let an AI agent buy my pants
- 46:26 Learning new topics with an LLM tutor
- 49:20 Would teenage me have learned anything?
- 53:32 Is it still making things up?
- 56:49 Hallucination or mistake? Ask for citations
- 59:23 Messy government and FDA data
- 1:02:35 Training your own model, and good-enough AI
Mentioned in this episode: