THE SECOND TIME YOU FIX AN AI, YOU'RE FIXING THE SYSTEM.
Proton's internal AI engineering manifesto has become a public argument about what software engineers should do when coding agents become part of the workforce. On October 29, 404 Media reported that Proton's CTO Bart Butler said AI wrote about 90% of the company's code in October. On October 31, Business Insider reported on the reaction to Butler's memo, which recasts engineers as leaders of models and agents.
These are reported statements, not independently audited evidence of code quality or productivity. 404 Media also documented employees' concerns about security, the pace of change, and workplace pressure.
The forgotten correction
Suppose an agent makes a mistake. A human fixes it. The agent produces a better result. Then the next session begins, and the same failure happens again.
That is not simply a model-quality problem. It is a continuity problem. The correction has not become a durable part of the workflow.
**IF THE CORRECTION DISAPPEARS AFTER THE CHAT, THE ORGANIZATION DIDN'T LEARN.**
A useful agent system should preserve what failed, under which conditions, which evidence established the failure, and which intervention worked. It should distinguish a temporary failure from an impossible action. It should also know when changing conditions justify reopening a previously unsuccessful path.
The Trillsverse connection
John Brajer's Possibility Reserve research includes Conditional Failure Memory, Graded Dormancy, Causal Reactivation, and Path-Expansion Search. Conditional Failure Memory attaches a failure to the conditions under which it occurred rather than permanently discarding a possibility. Causal Reactivation asks which changes should bring a dormant possibility back into consideration.
These ideas connect to a practical engineering question: after an agent fails, what information must survive for the next attempt to be meaningfully better?
The goal is not to make every model remember everything. It is to preserve relevant state and evidence in a system that can retrieve, verify, and act on it. Sensitive data, permissions, and access boundaries still matter.
Architecture includes accountability
An agent team still needs testing, security review appropriate to the risk, provenance, and human responsibility. Replacing manual work with automated repetition does not create a reliable system; nor does measuring how many lines AI wrote tell us whether those lines should ship.
John Brajer develops Trillsverse architecture. John B creates the music. John Braj carries personality, media, and cultural conversation. Trillsverse is the reality update.
**THE SECOND TIME YOU FIX AN AI, YOU'RE FIXING THE SYSTEM.**
What's one correction you've had to give an AI more than once that should already be part of the system?
Follow the work
- Conditional Failure Memory: https://reallythatmagazine.com/article/conditional-failure-memory-failure-belongs-to-conditions
- ReallyThat Systems: https://reallythatmagazine.com/systems
- Trillsverse Gate: https://trillsverse.com/
- 404 Media, October 29: https://www.404media.co/manifestos-fear-and-loser-thinking-inside-protons-ai-reckoning/
- Business Insider, October 31: https://www.businessinsider.com/protons-cto-ai-manifesto-software-engineers-coding-2026-10