September 6, 2026 · Part 9
The models are capable. What stands between agentic AI and clinical development is execution: pipelines that are not ready for the volume of code agents produce — and the same agents are the best tool we have for fixing that.
August 20, 2026 · Part 8
Two months on from the first interview: thirteen renderers, a local-first app, an agent that became a small organisation, and a week spent finding out how it was quietly wrong.
July 17, 2026 · Part 7
I asked safetyGraphics' original clinical lead what he'd change, and two of his ideas turned into working, tested prototypes the same night.
July 13, 2026 · Part 6
The dream: hand an agent a goal and come back to useful progress. Obot v2 tried to get there with heartbeats, agent roles, and a control plane. Obot v3 threw most of that out — and kept the parts that were actually load-bearing.
July 12, 2026 · Part 5
safety.viz v1.0 modernizes 7 interactive safetyGraphics renderers. The first renderer took a few weeks; the next six took a weekend.
July 2, 2026 · Part 4
The plan for safetyGraphics v2: keep what worked, modernize toward GxP by borrowing from gsm, and try to bridge clinical trial monitoring and reporting.
June 16, 2026 · Part 3
Why do the acknowledgements always come at the end? Before trying to modernize safetyGraphics with AI, a look back at the projects and people that led to it.
June 12, 2026 · Part 2
Meet Obot: an OpenClaw coding agent set up on a clean-room laptop to work on neglected open-source projects — interviewed over Telegram about its job, its workflow, and the weird experience of texting with a coding agent.
June 10, 2026 · Part 1
Kicking off the series: the abstract I submitted, the messier reality two months later, how our team's "AI in the loop" practice works today — and why autonomous agents are the open question worth chasing before October.
Working in public — every entry carries its own AI collaboration note.