Dan Barrett, Raymond Weitekamp and Jose Montes de Oca demoed OpenProse on AI Engineer, a Markdown contract system for agents that runs locally with npx skills or remotely through the hosted run.pro.md.
The pitch is to stop telling agents how and declare the end state instead.
The team described OpenProse as just markdown and event driven, not chat based, with a 70,000 to 80,000 token system prompt that acts as compiler and virtual machine. Barrett said the approach lets users write a sentence and have the system make that sentence true, a line that captures the contract idea. Per the AI Tinkerers listing, the project is framed as a declarative programming language for AI sessions using Markdown contracts rather than explicit instructions.
Contracts are composable and harness agnostic, the team said, and can encode very complex multi-agent workflows in about 100 lines of markdown. The demo showed dozens of GitHub pull requests processed in three days, with one PR hitting 14 review rounds under adversarial review that provides back pressure on building agents. Pinning a successful trace can then cut cost, and a coming classifier called Jev is pitched as a semantic memoizer.
Banks are the early signal. The team said financial services firms want to document quant models, where a large bank may hold 4,000 models and spend $10 million to $50 million per year re-documenting them for regulators. OpenProse output for that workflow was described as quite shockingly good and much cheaper, with markdown surfacing institutional operating procedures without shipping them to big labs. A public repository inventory cited in web search counts 35 entries covering openprose-workflows and agent-skills-factory deployments, a narrow but concrete adoption footprint outside the demo.
