Software & repo reviews
The rule for this section. Only software that actually runs here goes in this section, long enough to say something meaningful about it. Numbers are re-measured on the day of publication, not copied from a README. Bugs I encountered myself are included — with the cause and the fix, not just the complaint. Open source also means: no affiliate links, only a direct link to the source code.
The reviews
DeepSeek's agent framework, eleven days old when I tested it and already downloaded 648,000 times. Everything is a plugin turns out to be literally true: 129 plugins in one list, with per plugin visible which layer modified it. I ran into two blocking bugs, one of which made every tool use impossible unless you talk directly to DeepSeek — including the one-line fix.
A working environment in which an agent builds applications for you, running on the Workers runtime. The best find is in what's not allowed: gadgets run in a sandbox without fetch(), and everything that goes outside runs through gatekeepers that log what happened. Bulk approval after the fact, with rollback. I run a daily publishing pipeline on it — and ran into an agent that delivered a gadget that only pretended to work.
Why two opposing designs side by side
These two projects solve the same problem in opposite ways, and that's what makes them instructive together. DSH makes every component replaceable and leaves safety up to you: you can rewire the model layer in eleven lines of YAML, but nothing stops you if you do something dumb. Cloudflare OS does the opposite: safety is baked into the architecture, and building is left to the agent.
In my own setup, both run, each for what they're good at. What underlies that — a router layer that makes models interchangeable and tracks usage — is described in Smartly outsourcing AI tasks via a proxy and routing. If you want to determine for yourself whether such an agent does what it claims instead of assuming it, start with how to evaluate an AI agent.
What's coming next
On the list are a comparison of the three agent frameworks on the same task, and a piece about what happens if you let a scheduled agent task run unsupervised for a week. Both require measurement rather than a first impression, so they'll only appear once there are numbers. What comes in daily about updates on the software above, we collect in the newsletter.