A short one – heads down all week and a bit up against it.

"Be concise" doesn't work: it treats verbose replies as a compression problem, and the model either squeezes too hard (fragments, arrow chains) or decides this reply is important and writes long anyway. What worked was changing what a reply is: the Slack message a trusted colleague would send after going away and doing the work. I asked the same question under both: 400 words before, 202 after, and nothing missing. The full instruction file is in the post, free to copy. Blog post

Downgraded my office to a hot desk: most of the systems are built now and the agents do the doing, so I stopped needing a proper work setup. I've been enjoying floating between offices and blending working with living, instead of being chained to one desk all week. One change I need to make soon: the agents still run on my laptop. Moving that to the cloud means the work runs at any time, and I connect in remotely – laptop or phone.

On my radar

Steve Yegge's future-of-agents essay: keep seeing this one everywhere, with reactions split between "mostly right" and "losing it". The claim to take seriously: human code review ends, and agent harnesses replace CI/CD (the automated test-and-release pipeline). Essay

The harness is the bottleneck: a week of rate-limit complaints and people switching providers. This piece argues the model was never the constraint – DeepSeek's budget coding model holds up as a daily driver at ~$0.19 a run, once the harness (the tooling wrapped around the model) is tuned. Blog post

A 150M-parameter model on ARC-AGI: Pathway's tiny model scores 29.5% on the reasoning benchmark at $0.0007 per task – GPT 5.6 Luna scores 34.2% at roughly 11x the price. The cheap end is catching up fast. Pathway