From blockchain to AI agents

A personal project of mine started as a blockchain platform. It turned into a team of AI agents, and that change taught me more than the build did.

By day I work in growth. Outside of work, I’ve spent the last few months building FinChat.

Version 1 had everything: tokenomics, smart governance, and an AI system monitoring every conversation, all anchored on the blockchain. It looked great on a whiteboard.

But the part people actually cared about was the AI, the piece I’d built as a side feature. So I followed that signal:

• The token economy became AI budgets that keep costs predictable.
• The AI monitor became six specialist agents covering markets, careers, research, portfolio tracking, and one that diagnoses the others when they fail.
• The blockchain stayed, but only as a tamper-proof audit trail.

The hardest lesson? Evaluation.

My scheduled agents were marked “successful” while delivering error messages as their reports. When I checked what they actually delivered, one agent was failing 73% of the time.

Building AI is getting easier. Knowing whether it actually worked is the hard part. I think that’s why evals have become such a big focus for AI teams, and why defining “success” is becoming a core product skill.