Tag
Posts tagged “ai-development”
The Boundary I Stated Out Loud Had a Shelf Life
From 14 August, Claude Code runs in auto mode by default and a classifier approves tool calls instead of me. Anthropic has the numbers to justify it and they are not close. The part that changed how I work is buried in the docs: a boundary you state in conversation is stored in the transcript, and the transcript gets compacted
Aug 12, 2026 · Robert Bojor
The Second Agent Is Not a Second Source
Two of my research agents returned the same claim citing the same URL, and the URL did not say it. I caught that and felt clever. Then I found out the claim was true, my correction had thrown away a real fact, and the independence I thought I was buying had already leaked out through three channels I never looked at
Aug 10, 2026 · Robert Bojor
Isolation was never on the list
Anthropic documents git worktrees as a way to stop parallel sessions colliding. Its page on choosing isolation for your threat model does not list them at all. Its changelog has six fixes in twenty-eight days stopping worktree-isolated agents from reaching the main checkout. All three of those are true
Aug 5, 2026 · Robert Bojor
The Protocol Stopped Remembering You
MCP shipped its 2026-07-28 spec and deleted the handshake, the session, and the server's ability to call back into your model. The operational case for it is good. The bill lands on everyone who now has to hold the state themselves
Jul 29, 2026 · Robert Bojor
The Most Careful Number in the Room
Microsoft measured what AI coding agents do to output and got a real number: 24% more merged pull requests, with a confidence interval and a placebo test. Then it made a decision about those exact tools that the number played no part in
Jul 22, 2026 · Robert Bojor
MCP Minus the Hype: What Wiring Models to Tools Taught Me
Model Context Protocol from a developer who shipped with it: what it fixes, the security part the tutorials skip, and when to skip it entirely.
Jul 9, 2026 · Robert Bojor