MARIO.
Multi-Agent Runtime for Intelligent Output
The agentic harness I built for Woven Health Collective, used across email, websites, congress kiosks and internal-tool builds.
It starts every build already onboarded.
Woven’s SOPs, development standards, client requirements and compliance rules load before a session starts. AI is a brilliant new hire who doesn’t know how the company works; this is its onboarding.
It plans first, then stops for a person.
It states the scope and how it will reach the goal. Nothing is built from a plan a person hasn’t read.
Specialist agents build, review and harden the work.
They build the approved plan the way the development team would. An independent AI reviewer checks it, then automated quality and security checks run.
A person approves again before anything ships.
Issues are flagged, never resolved silently. Approved copy and design are preserved.
Every correction becomes a lesson for the next build.
Reviewed lessons go into a shared library that every plan reads first. Feedback from client and MLR review joins the same loop.
At a gate, a reviewer can change the plan as well as approve it.
The plan is held with its tasks, criteria and boundaries until a reviewer adjusts or approves it.
One harness works across any channel and any model.
Every channel comes out of the same process. It isn’t tied to one AI model, so models can check each other’s work.
It gets better with every project it runs.
- 23lifecycle skills
- 11tools in its MCP server
- 7stacks
- 750+reviewed lessons
- 50+projects
- ~60%less developer time and costobserved across projects
Hear how it works from the people who built it.
Cayce Richardson of Woven Health Collective talks with Tim O’Sullivan, EVP of Technology Innovation, and me.