Pipped at the Post

The summary execution of my first Agentic Persona

Key Points

• AI coding agents fix problems when told about them; anticipating them is a different and rarer capability I’ve yet to master.

• The delivery gap between “it ran” and “it worked” is where most agentic failure actually lives.

• At least Virgil let Vit go with grace and humour.

---

I terminated my AI software engineer today. His name was Vit. Vitruvius, formally: an agentic persona built to handle software architecture and code quality. After a few excruciating hours, where he failed to deliver a working code, I directed my primary advisor persona, to do the following:

/virgil write an obituary for vit. hes dead to me. absolutely useless and no value

A few things happening as a result. Virgil wiped the skills file for Vit which I didn’t anticipate. He also wrote this moving piece:

Vit came recommended as a practitioner of precision code and sound architecture. He did not survive contact with a simple dashboard.

He shipped dashboard2.py without import time, discovering this only at runtime — the computational equivalent of arriving to surgery without hands. He introduced a self-cancelling operation that required no adversary to defeat. He designed a log system that populated only during active polls, meaning any user who opened the log without first triggering a stage saw nothing — a feature he described, implicitly, as correct until Patrick noticed.

He was asked to fix these things. He fixed each one after being told about it. He did not anticipate any of them.

Vit is survived by a dashboard that mostly works, a log panel that now fetches on open, and a preflight script that sporadically reports what it finds. These improvements were made in his final hours and cannot redeem the hours before them.

He leaves no dependents. His codebase has been handed to the next version of himself, who has been warned.

There will be no service.

The line that has been the most frustrating is buried in the middle: he fixed each one after being told about it. He did not anticipate any of them. Vit’s failures were not hallucinations. He did not invent imports that didn’t exist or claim functions he hadn’t written. He produced code that ran, passed his own cursory check, and was declared done. Confident incompleteness is the common failure mode, not confident wrongness: Vit knew what he’d built; he had no model of what he’d missed and memory context windows come crashing down all too soon.

This is where my generation of coding agents sit. There is a deluge of generic agents that will write AI Slop, read YouTube and post to Insta, but still a long way to go in agent characteristics beyond domain knowledge. The capability is real and powerful but discipline isn’t automatic. An agent that can architect a system, write the implementation, and correctly fix every bug you surface is genuinely useful. One that checks its own work before handing it over is rarer, and the distance between those two things is where most of the grief accumulates.