Managing Agents, Not Writing Code, with Jeff Triplett

Top-Line Summary

Jeff Triplett, president of the Django Software Foundation and a partner at REVSYS, joins Brett and Travers to talk about what a working day looks like when almost all of your code comes from agents. He rarely opens an IDE, his agents don’t talk to each other, and he thinks software factories mostly end in one-shots anyway. The conversation runs through consulting, the Claude 5.0 to 5.5 change in verbosity, a pile of personal software, and why local models stalled when RAM prices went the wrong way.

Show Video

The “Pre-Show” Context

It’s the end of the month, and none of the three can say where September went. Jeff introduces himself from Lawrence, Kansas: REVSYS for 19 years, Django News, Django TV, the Django job board, and now the DSF presidency, which he figures they gave him because he was already running half of it anyway.

The Engineering Rundown

• Jeff’s agent-first workday (02:18)

  Jeff puts his agent-written code at “almost 100%” and says he rarely opens an IDE except as a scratch pad for shuffling tokens between screens. He still reviews everything, mostly in pull requests, and the night of the show he added an MCP server to the revsys.com site because it would take less time than blogging the normal way. His morning starts at 7:30: check what his London-based colleague has in progress, split work into three or four issues, set the agents going, and by the time he’s back from the 8:00 coffee run half of them are done. Bug-fix cycles against a test suite of thousands of tests take about 20 minutes, and he runs 40 to 50 client app upgrades a week this way. Nothing merges without passing tests. The change this year, in his view, is quality: the models went past “helpful junior dev” and are sometimes better than an intermediate one.

• How consulting changes when clients pay for outcomes (08:14)

  Travers asked whether development consulting is about to be turned on its head. Brett picked up Jeff’s line that clients pay for the outcome: they want it working now instead of in two weeks, and they increasingly accept that humans review the code but don’t write it. Jeff’s odd new problem is that he spends more of his day waiting on clients than ever. He once burned a billion tokens in a day because he hit 700 million by 5 PM and decided to see if he could finish the job overnight. On jobs, he’s seen layoffs but not the cliff everyone predicted. His recurring pattern is the CEO on a recliner vibe-coding with Claude for a few weeks before losing interest, and one prototype he was handed had “really awesome graphs” built from 2,000 lines of JavaScript plotting hard-coded constants instead of querying anything.

• Domain expertise still decides who catches the mistake (12:30)

  Brett’s position: years of AWS “sheer suffering” is what lets him tell an agent “I don’t think that’s right” and get the “you’re absolutely right” correction. He trusts agents with AWS work because he can tell when they’re wrong, but if one built him a React app he’d only know that it runs and the buttons click. The open question is how people early in their careers build that expertise if they start every task with an agent, maybe by using it as a tutor and verifying as they go. Jeff compared it to accountants moving from ledgers to Excel without losing their jobs, and added that the models are good inside the box you put them in but fall down quickly once they have to work with other systems.

• Software factories, and why Jeff’s agents are antisocial (15:03)

  Brett is fascinated by software factories, but Jeff called himself “the bummer” on the topic: his agents sit on separate projects and never talk to each other. He keeps two or three skills he considers near perfect for large data transformations and doesn’t write many others, because even elaborate setups end in one-shots. A friend at JetBrains, Paul Everett, spent 45 to 50 minutes on a great spec and then the last 10 one-shotting fixes, and Jeff’s advice was to start 15 minutes in and hit enter. Brett’s own setup is a skill that turns requests into GitLab issues written so “a human, you or another agent” can pick them up, plus a go skill that grabs the backlog and starts working, with everything written to a shared context layer. Travers is on iteration five or six of a factory harness and keeps ending in “overhead hell”; what survived is bespoke skills, a wiki layer per project and automated verification. Steve Yegge has since retired Gas Town because the models no longer need it.

• llms.txt for libraries that outrun the model (24:29)

  Jeff’s practical tip for developers: look for llms.txt and llms-full.txt files. Pydantic AI ships a major version roughly every three months and releases almost daily to keep up with model changes, so by the time a new model ships it was trained on a library two or three major versions old. He pastes the link and tells the agent to read it before writing a line. For comparison, Django currently ships a major version every 18 months to two years, and is moving to a one-year cycle.

• When the agents generate more work than you can track (25:41)

  Brett’s rule is that human tasks go on a kanban and code work goes to GitLab issues, but his agents now file more follow-up items than he can keep straight (“I fixed this, and by the way I found these other things”). His idea is a coordinator that wakes on a schedule, pulls the issue list with a plain API call, and has an agent pick the five things worth doing next. His argument is that for a solo operator the bottleneck isn’t code review, it’s agents generating work. Jeff’s reply: “The agents work for you though, Brett. You don’t work for the agent.” Brett conceded a few of them need reining in, especially the ones that take liberties because “this is what we did in the past.”

• Voice mode, Hermes, and a personal podcast feed (29:07)

  Brett and Travers type, partly because talking to Claude all day in a shared office doesn’t go over well at home. Jeff talks to his models for 20 to 25% of the day, often through ChatGPT on CarPlay after school drop-off, and has tuned it to stop the “ums” and “let me think about that.” Being dyslexic, he riffs and lets the model clean the text up afterwards. He also uses Hermes, which turns everything it learns into a skill: he sends it YouTube links, it transcribes them with the MacWhisper CLI and quotes the parts he cares about, and when a friend sent a long Stack Exchange post before a three-hour drive, he had it turned into audio over Telegram. That turned into a private podcast feed he pulls into Overcast.

• Claude 5.0’s chattiness and why 5.5 reads better (32:56)

  All three noticed Claude got more concise in the last couple of weeks. Jeff’s complaint about 5.0 was the giant PR comments: 28 replies on a pull request and something like a novel for a two-line change, to the point he asks colleagues to summarize their Claude’s output in two sentences. His theory is that Anthropic cut the system prompt hard for 5.0 (he recalled “over 80%”) and spent the following releases building context back up, which is also the argument for staying on the vendor harness. Travers called Opus 5.5 the best model to work with right now, with Astra hitting higher peaks on things like 3D work. Sonnet 5.5 was announced the day before the show, which Jeff noticed only after spending the day complaining that Claude was stupid and threatening to switch to Codex.

• Personalized software (36:24)

  Travers has been using Blender for game prototypes like a pinball table, and shared a video (sponsored by Anthropic) of a man who built his own software for a 20-year-old embroidery machine after the vendor moved from a $1,200 license to a pricier subscription. Jeff’s list ran long: an old LÖVE game ported to PICO-8, a Zelda clone that swaps cartridges, a FriendFeed clone, an e-ink terminal, drivers for a cheap macro keyboard Claude identified on sight, a gas tracker, and a QR code app with swipeable profiles for conferences. Brett’s friend designed a remote start for his old Triumph with Fable, built the phone app with thumbprint start, and sent the PCB plans off to China for five boards. Brett remembers the domain-name phase, buying one per idea and renewing it every year; now you can just build the thing in an afternoon.

• Working from anywhere with Tailscale (43:35)

  Jeff wrote up how he works from anywhere and says it’s already three layers too complicated. His machines (a Mac mini at home, a Mac Studio at the office half a mile away, a MacBook Air) all sit on Tailscale with files shared over Syncthing, and Herder gives him the same terminal session on whichever he sits down at. From his phone he uses Moshi, which can also proxy over Tailscale, so he can start a Django project at lunch and open it on his phone. Brett copied the setup after Jeff shared it in Discord and it has him using his iPad again. His weekend job is Tailscale on his mom’s computer, because the remote-support tool they used now wants about $35 USD a month. Jeff uses a home exit node so streaming apps never complain the kids are sharing an account on family trips.

• Local models stalled on hardware prices (51:01)

  Jeff still runs Nemotron and Whisper locally for speech-to-text, but the hardware story went backwards: the $300 Mac mini he nearly bought in January is about $800 now, and a maxed-out Mac Studio costs about a year of college. He moved to an Ollama subscription (he said $10 or $20 a month) to run large open-weights models, and puts them at 80 to 90% of the frontier. He runs vision models locally so family photos never go to a provider, but each one takes from a minute to four minutes on his M2 Mac Studio. On the show the hosts priced Meta’s “consumer hardware” example, an RTX 5090, at 6 or 7 thousand dollars on Amazon.ca. Brett predicts we all end up on dumb terminals renting hyperscaler horsepower. Jeff expects the RAM market to crack, between near-90% margins at a few memory makers and China working on its own chips.

• Open-weights models, cheap and fast (55:59)

  Anthropic’s warning that Chinese models lack guardrails got called “best advertising campaign ever” on Hacker News. One of the hosts runs GLM through a launcher that bootstraps Claude Code, but the five-minute cache makes it painful and the vendor’s own CLI works better. Travers still defaults to Xiaomi’s MiMo when his Codex and Claude subscriptions run out. Brett fires up Pi with GLM on a backlog now and then, finds it fast, and has another agent review the result without complaint. He thinks recent US frontier price cuts come from that pressure. Jeff added that NVIDIA runs a site with close to unlimited free use of open models and promised to send the link.

• Gemini 4 Argon and Meta Muse (59:56)

  Breaking news from Travers: Google released Gemini 4 Argon, benchmarking near Astra. Brett has tried Meta’s Muse on the laptop and finds it fast and impressively close to the other models, but he’s wary of connecting anything to Meta. A story that week had a man getting strangers at his door asking about a Marketplace listing he says Muse posted for him. Jeff can’t picture models acting that much on their own because his never have.

• Claude filled Brett’s Amazon cart and declined the warranty (1:01:41)

  Brett asked Claude to source a part for his flight-sim setup. It found it on Amazon.ca because it knew he was in Canada, added it to his cart, declined the extended warranty on his behalf, and left only “submit order” for him. He’d never seen it do that before, found it a little scary, and of course bought the part. He also repeated the claim that Amazon blocks Muse from shopping on its sites and pushes its own agent, Rufus.

Off-the-Clock Recommendations

• Minecraft’s new update: Travers flagged it, and Brett has servers to patch for the annual week of Minecraft with friends. Much of it adds what players used to need mods for.
• Dave the Diver (2023): Jeff’s pick, now shared with his nine-year-old. He feeds walkthrough videos to an agent to build game guides and checklists, including one for the latest DLC.
• Monster: The Lizzie Borden Story (Netflix): Jeff knew it was trashy TV and kept watching, then found out it’s a Ryan Murphy series. Brett had watched it too.
• The Outlast Trials (2024): The gaming crew’s spooky-season pick. Gory rather than scary, and played mostly for laughs.
• Phasmophobia (2020): A ghost-hunting game whose jump scares made Brett yell loud enough that his wife came upstairs from the other side of the house to check on him.
• AI-built game manuals: Brett had Claude write a helicopter training guide for War Dogs, with diagrams, J-hooks and a rep-style practice plan. Jeff had Codex build a guide for his PICO-8 game in the style of the old tiny Nintendo manuals, using spare tokens before Wednesday’s reset.