a weird inversion with LLMs is the models improve faster than the tinkerers when i see people with custom workflows and setups they're all addressing problems that don't exist anymore the person naively using vanilla codex is more likely to be experiencing state of the art

The most important technology problem of the next 5 years: Creating a broadly accepted standard for agents to identify themselves to service providers, so that providers can adjust the way they interface with clients (as compared to human-based traffic). In the interim (i.e., for the next 6 months), there will be a cat-and-mouse game that agents will play -- to circumvent detection and maintain their product's utility during this growth phase. However, this will only be a stopgap and it will not be the terminal state of the world.

Every developer needs an AI shed: An always-on machine running on their tailscale network where the majority of their herdr agents are running.

Asked a friend, who works in a successful business and is heavily using AI for building things, why he would use a SaaS vs building it himself. This was his answer. Sure, you have to pick the product, so that's your liability, but that's why you pick SaaS with a good reputation.

i still don't get this position if you get an ai generated PR you can just... close it? have a review bot that automatically flags and closes slop PRs? or put instructions in agents.md that AI generated PRs will follow?

Serving a $DASH order through an AI agent costs $0.02 to $0.20 cents of compute, against $1.70 of contribution profit per order. The napkin math: Unit economics, from the Q2-26 10-Q: • Average order: ~$34 • Revenue: ~$4.60 (13.5% take) • Contribution profit: ~$1.70 (5.0% of order value) • Adj. EBITDA: ~$0.95 (2.8%) Everything below is measured against that $1.70 contribution profit. Moving forward, an agent doesn't make one call. It browses, compares and checks out, and every step re-reads its context. At market prices for frontier-class open models (~$1.30 per million input tokens, ~$0.10 cached), a session costs (est.): • Reorder "my usual", ~45k tokens: ~$0.02 • New discovery order, ~135k tokens: ~$0.06 That is 40x to 120x the compute behind a Google search. $GOOGL earns ~4¢ per search and spends ~0.05¢ of compute on it (est.), and its largest variable cost is the ~0.7¢ of traffic acquisition it pays Apple, Samsung and others. Advertisers pay per click, so they carry the searches that don't convert. An agent paid on a take rate carries a different equation. For DoorDash that means the failed sessions land on the completed orders. If 80% of reorder sessions convert, compute is ~2.4¢ per completed order, and if 30% of discovery sessions convert, it is ~20¢. Both conversion rates are my assumptions, not disclosures. Contribution profit per order: • App order: $1.70 • Agentic reorder: ~$1.68 • Agentic discovery: ~$1.50 Compute takes 1.4% of the margin on a reorder and 12% on discovery. Those two ratios are also the break-even. An agent order that would have happened in the app anyway carries the compute as pure added cost, so the agent pays for itself once 1.4% of its reorders, or 12% of its discovery orders, are orders DoorDash would not otherwise have had. What absorbs the compute is the take rate. The same $34 discovery order on a payments-style 0.3% fee would need 58% of sessions to convert to cover it, and at DoorDash's 13.5% it needs 1.3%. Merchants of record on marketplace takes ($DASH, $BKNG, $EXPE, $AMZN) can add an agent for cents per order, while a third-party agent on a thin fee is confined to repeat and big-ticket purchases until merchants can no longer refuse it. So the exposure sits on the revenue line. Ads reached a ~$1B annualized run-rate in 2025, about 7% of DoorDash's $13.7B of revenue, which is 26¢ to 33¢ of the $4.60 it earns per order (est.). They are sold in the app feed, and an agent order never opens the feed. The agent's costliest order burns 20¢ of compute, and the average order carries ~30¢ of ad revenue it doesn't show. But there's a scenario Doordash put sponsored options for the agents, ensuring those ads were read by the agent, delivering better ROAS for advertisers. The line to watch as agent orders scale is net revenue margin, 13.5% in Q2-26, since that is where displaced ads would show. I'm wrong on compute if contribution profit falls below 4.5% of order value with management citing AI serving costs.

Lunch, as it turns out, is strangely a very difficult problem to solve (exponentially harder as the number of teammates grows). It's been fun to see how different companies and teammates are using our CLI to make lunch an easier thing to do. Now we've built it into an MCP to get companies all things food and more!