Sponsored by
Two numbers this week have nothing to do with each other and still rhyme: one tool swap can make an agent 25 times faster, and on the other end of the spectrum, one billing change can make running a model up to 20 times more expensive. Same family of coincidence, two completely unrelated stories. I'll start with the one I clocked myself.
This week
The default ritual for an agent driving a browser: bolt on click/navigate/screenshot tools and let it drive like your grandpa renting a car in a country where he can't read the road signs. One pixel at a time, a screenshot after every move, just to check nothing's on fire. It works. But every micro-click is a full round trip to the model: "here's what I see, here's what I'll do next, wait for permission, do it, screenshot again, repeat." I run Claude Code against 30-plus repos a day, so "does the deploy actually work" isn't a rare event for me, it's basically a heartbeat.
I tested one page, four checks, nothing exotic: navigate to a live site, grab the page title, extract the nav links, take a screenshot. With a standard click-driven MCP tool: 4 tool calls, roughly 40 seconds, ~135 KB dumped into context (full page text plus a base64 screenshot, whether I asked for it or not). With a CLI that scripts the browser instead of driving it by eye: 1 tool call, roughly 1.5 seconds, ~1 KB of data, the exact bits I chose to log. Same site, same four actions. The whole gap comes down to one line: most of the internet doesn't need improvisation, it needs you to show up and read the label.
Tomorrow morning, before you plug an agent into a browser again: ask yourself whether the task genuinely needs to improvise at every step, or whether you already know the path. If you know the path (checking a deploy, scraping a known page, extracting links), a direct Playwright script beats a click-by-click MCP tool on both time and tokens, in a ratio that gets close to 25x.
The AI assistant for admin tasks
Scheduling meetings isn't why you get out of bed. Filing expenses isn't your calling. So hand it all to Catch — the AI admin that takes the tedious work off your plate and gets it done.
Just ask. Catch handles the back-and-forth on scheduling, flights, restaurants, follow-ups, vendors, clients and more — then comes back with it sorted, in moments.
No guesswork. No dropped balls. What's yours stays yours, private by default and reliable by design.
Nobody became a leader to book their own flights. Meet your admin savior and go from swamped to sorted.
Get started free at catchagent.ai — and start spending your time on the things only you can do.
In the news
Here's the other half of that opening parallel. As of July 12, free-tier inclusion of Claude Fable 5 in Pro/Max/Team plans is over (the cutoff had already slipped once, from July 7 to July 12). Access now runs on metered usage credits billed at API rates: $10 per million input tokens, $50 per million output, twice Opus 4.8's rate. Anthropic hasn't published an official conversion between the included weekly quota and a dollar figure. One community estimate, not an Anthropic number, puts sustained Fable 5 usage on credits at 10 to 20 times the price of a premium subscription to keep the same pace of work.
This week on the blog
- I Made My AI Agent Stop Clicking Around Like a Confused Intern and It's 25x Faster Now: the full benchmark behind this week's edito, numbers table included.
- The one-person company stack I'd run in 2026, all self-hosted: the full stack for running a one-person company without depending on a third-party SaaS.
- What a self-hosted AI agent actually costs per month, real numbers: the actual bill, line by line, not a vague estimate.
Revenue Stack: July 2026
The lesson that applies to your content or monetization stack too: stop eyeballing what you could script once and be done with. The time saved on every micro-check adds up faster than you'd think across a week of 30 repos.
How does your AI agent drive a browser today?
Phil
PS: reply and tell me the last task where you let an agent improvise when you already knew the path.
Back to that 1.5 seconds. Click-by-click driving isn't a bad idea, it's the right instinct for a genuinely unpredictable UI. The problem is using it by default on the 90% of tasks where you already know every step in advance.
You're receiving this because you signed up at rentierdigital.beehiiv.com.