Jeff Triplett, president of the Django Software Foundation and partner at REVSYS, writes almost none of his code by hand. He manages agents instead, running 40 to 50 client app upgrades a week and 20-minute bug-fix cycles against thousands of tests. Brett and Travers ask him how AI coding agents change development consulting when clients pay for outcomes, and why domain expertise still decides who catches the agent's mistakes. Jeff explains why his agents never talk to each other and why software factories end in one-shots, and gives a practical tip about llms.txt files for fast-moving Python libraries like Pydantic AI. Brett describes a different bottleneck for solo operators: agents that file more work than he can track. The three compare Claude 5.0's chatty pull request comments with 5.5, swap stories about personalized software built in an afternoon, and walk through a work-from-anywhere setup with Tailscale, Herder and Moshi. Then local LLMs, rising RAM and GPU prices, cheap open-weights models like GLM and MiMo, and the time Claude filled Brett's Amazon cart and declined the warranty for him.

Subscribe for weekly AWS and AI engineering talk, and tell us in the comments how many of your agents' issues you actually read.

Chapters:
00:00 Where did September go? Meet Jeff Triplett
02:18 Jeff's agent-first workday
08:14 Consulting when clients pay for outcomes
12:30 Domain expertise and catching agent mistakes
15:03 Software factories and antisocial agents
24:29 llms.txt for fast-moving libraries
25:41 When agents generate more work than you can track
29:07 Voice mode, Hermes and a personal podcast feed
32:56 Claude 5.0 verbosity vs 5.5
36:24 Personalized software
43:35 Working from anywhere: Tailscale, Herder, Moshi
51:01 Local models and hardware prices
55:59 Open-weights models: GLM, MiMo, price cuts
59:56 Gemini 4 Argon and Meta Muse
1:01:41 Claude filled the Amazon cart
1:03:13 Recommendations