Prisma's engineering team reflects on their evolving approach to agentic software development. They started with a heavily opinionated 'Drive' process — structured skills, multi-agent workflows, and explicit artifacts — which proved agents could ship large-scale changes with high quality. Over time, the verbose process became a maintenance burden as models improved, making baked-in assumptions obsolete. The team is now moving toward leaner skills that define contracts (inputs/outputs) rather than narrating processes, using frontier models for planning and cheaper models for execution. Key durable principles include clear specs, explicit plans, testable acceptance criteria, review loops, and human judgment ownership — while keeping the workflow itself adaptable as tooling evolves.

13m read timeFrom prisma.io
Post cover image
Table of contents
Why we started with an opinionated processWhat the opinionated path provedWhere the opinionated path started to cost usThe lean path: skills as interfacesFrontier planning, cheaper executionCost, without token maxingDon't overbuild around temporary limitationsWhat leadership should standardiseProcess as a bridge, not a destination
659 Impressions