What's your PR pace? I've been thinking about development velocity for a while, because nearly every problem I need to solve comes back to building software well. Lately it's felt like I've found a flow so I wanted to see whether that was real. PR count treats a 10-line PR the same as a 1,000-line one, so I designed a metric that ‘weighs’ each PR by its size and had the agents build a CLI for it. It was real. From June to October my PR count doubled, but size-weighted PRs went up about 7×. You can try it by running npx pr-pace pace in any repo you merge PRs into. It's my first npm package and a proper write-up is coming. GitHub

The PM agent runs itself now: I got to the point where I was approving nearly every recommendation my agents make, so my product manager agent now makes any call that can be undone (a two-way door). It launches build sessions that report back to it and merges their PRs once they pass a set of criteria, with anything that can't be undone still coming to me. The most useful thing stress-testing turned up: a call only counts as reversible if someone would notice it was wrong. It needed (and still needs) fine-tuning, but it's magic to watch. Write-up coming.

Outbound, 0 to 1: I was so heads down on the outbound engine (the agent pipeline that finds and researches prospects) that I forgot last week’s build notes. It now has a steady stream of people to email at accounts that genuinely fit. That took far longer than expected (and it'll never be finished I guess) but the 0 to 1 is done.

Memory debt: A founder chat this week got onto agents turning one edit into a permanent rule. My weekly review has an agent sweep every agent's saved memory and test each entry for two things: is it stale (wrong, or pointing at something that's gone), and is it dead weight (still true, but not worth loading into every session)?

On my radar

Consumer agents: Keep hearing about Grok Bot and Instinct handling people's life admin with the sticking point being trusting whoever holds your data. Anthropic's version (Claude Duet) opened a US waitlist yesterday. Announcement

OpenCode breach: A breach at OpenCode exposed the AI keys users had stored with it and some of them were hit with charges they hadn't made. Worth checking which tools hold your keys and whether each one has a spending limit. News

An obituary for programming: The fortnight's longest argument was about what's left for developers. Simon Willison's short post is about what people who loved the puzzle of coding are losing. Blog post