Problem

I wanted to have my agents build things for me, because I have occasional ideas that I (and possibly only I) think are cool and fun, and even useful, and I don't end up ever building them because time, especially programming time, goes to other priorities, such as work. I tried to have an agent build some stuff, like an awesome 3D dice probability calculator (since I like role playing games), and I felt like a cartoon character trying to pack too much into a suitcase—something would always fall out and the whole thing might explode. Not to mention delay and repetition inherent in fixing things via agent converstaion.

‍

Steps to Solve

1. My colleague and I built a skill to set up an autonomous harness that my agent could run to develop against a spec (https://github.com/fulcradynamics/community-skills/tree/main/skills/fulcra-rapid-prototype)
2. Using that skill, I answered an extensive but fast series of questions about what I wanted it to build (fast because the agent didn't have to do anything between questions, and because each question was simple and quick for me to think about). This idea was based on Matt Pocock's grill-me skill (https://github.com/mattpocock/skills/tree/main/skills/productivity/grill-me)
3. I ran the harness a couple of times "manually" with my agent to fix set up and permissions issues
4. I let the agent loop the harness until it built my dice calculator. You can use it at https://dier-84b61b31701333b6.surge.sh/

‍

Solution

The harness cron job with separate generator and evaluator, a separate cron job to review each harness run, and a grueling spec conversation up front, all release the user (me) from having to walk my agent through every step while waiting for it along the way. They also give the agent a place for knowledge and progress surrounding the project (my Fulcra account), rather than having to reread all the code and rely on session history. That same permanent record of what's actually happening allows me to keep in touch with the agent's work without reading a labyrinth of text walls (here is the dashboard you can see https://dice-harness-218b71836adea41b.surge.sh/)

‍

Result

I am happy with the basic dice probability calculator. It has a 3D comparative roll-probability graph (which I want to improve since there is no good reason for it to be 3D right now), and it stores my rolls in browser storage so I cam come back to the static site anytime and have my history.

My favorite part is that the dice calculator worked the first time. Everything worked. When I tried before to just have an agent build this in one shot (without features like the roll history), nothing worked the first time, and while more things worked the third time, something else was broken the fourth time.

I think the skill and harness have a ways to go. I think the harness is too heavy, and should be faster to run with less built-in rigidity. The dice calculator was a small project (though I may bloat it by adding more milestones to the harness, especially 3D ones). I want to experiment with bigger deliverables to see just how light and effective a generic harness definition can be. That said, it worked.