Skip to content
00 · Overview P. 01

Field notes from building software with agents.

Biosphere Labs publishes field notes, experiments and working tools from AI-assisted software development, agentic engineering and agentic software factories. Underneath it is one open question: whether a very small number of minds, directing many agents, can now build better, faster and cheaper than a large team that has to coordinate. The writing follows real builds to find out, including the ones that fail.

A system that needs to change all at once? Ask for a diagnostic →

Fig. 0.1
Live traffic is routed to the incumbent, which serves users. The same traffic runs through a shadow system on cloned data. Outputs are diffed and divergences classified as intended, defect or ambiguity. Live dataCloneCloned nightlyLive trafficRouterIncumbentServes 100% of usersResponse to usersShadowNew system · no usersΔIntendedDefectAmbiguityRuns in parallel until the cutover criterion is met12345In productionShadow
Drawing
Shadow run
Scale
NTS
Rev
A
Fig. 0.1. A shadow run. The existing system keeps serving users while a rebuilt one runs on a copy of the data, and every difference in their outputs is classified.
Rev. 1.0 · Oct 2026 01
01 · Writing P. 02

Featured

Essays

Why big rewrites may stop being incremental

For twenty years the safe way to change a large system has been one piece at a time. Agents change the arithmetic behind that advice. What the published evidence shows so far, and what it doesn't.

Published · 9 min read

Tutorials

Your Skill Only Works Because Claude Is Smart

I have a skill file that writes illustrated blog articles. Under Claude Code it behaves beautifully: it researches the topic, drafts prose in my voice, generates images, checks

Published · 20 min read

All writing →

Rev. 1.0 · Oct 2026 02
02 · Subscribe P. 03

New notes by email.

Each new piece as it's published. Nothing else.

Check your inbox to confirm.That didn’t go through. Please try again.