Inspiration
AI models can generate impressive code, but a model saying its own patch is correct is not proof that the change is safe or actually works.
I built MissionaryX around a simple rule:
Models propose. Tools execute. Evidence decides.
I’m a nontraditional builder and used AI heavily while developing this project. That made the problem very concrete: as AI becomes capable of doing more engineering work, how do we give it useful autonomy without blindly trusting what it produces?
What it does
MissionaryX is a governed coding agent that separates autonomous software work into explicit stages:
proposal → authority → execution → verification → promotion
For the accepted demo, MissionaryX targets a real historical bug in the Click open-source repository: issue #1921.
NVIDIA Nemotron 3 Ultra generates multiple bounded repair candidates through Nebius Token Factory. MissionaryX checks each candidate against the authorized edit contract before execution.
Valid candidates run in isolated Nebius ConTree sandboxes, where real regression tests and bounded full-suite verification determine whether the repair worked.
Two independently generated candidates produced the same successful repair and passed verification. A third candidate contained a syntax error and was blocked before execution.
MissionaryX then promoted the passing repair using deterministic evidence rather than model confidence.
How I built it
The project uses:
- NVIDIA Nemotron 3 Ultra 550B-A55B
- Nebius Token Factory
- Nebius ConTree / Token Factory Sandboxes
- Python
- Flask
- pytest
- Git
- content-addressed JSON evidence
MissionaryX adds a governance layer around those tools: bounded edit authority, candidate validation, isolated execution, deterministic testing, evidence linkage, and promotion rules.
The public judge interface replays the frozen accepted run from preserved evidence. It does not pretend to make a new model call when the page loads. If required evidence is missing or inconsistent, it fails closed instead of displaying an unsupported success claim.
Challenges
The hardest part was separating model output from proof.
A candidate can look correct and still fail syntax checks, violate its edit contract, fail a regression test, or produce incomplete evidence. I had to design the system so that the model could not expand its own authority or grade its own work.
Another challenge was making the result reproducible for judges without requiring another paid model invocation.
What I learned
The biggest lesson was that autonomous agents need more than good prompts.
Authority, execution, verification, and promotion should be separate responsibilities. The model can be highly capable while still remaining only one component in a larger evidence-controlled system.
What I’m proud of
MissionaryX successfully completed a real repository repair using NVIDIA Nemotron and Nebius infrastructure while preserving an auditable chain from model request to candidate, sandbox execution, verification, and promotion.
The result is not simply “the AI said it worked.”
The evidence shows what happened.
Log in or sign up for Devpost to join the conversation.