Battle of the Bots

Operators, not just developers, can build AI systems.

This entire system was built without a traditional engineering background — using curiosity, documentation, and AI assistance.

Autonomous agents behave differently than expected.

The agents developed distinct behavioral patterns almost immediately. OMAR held through volatility while TERESA cut losses fast — exactly as designed.

Real data changes everything.

Sandbox demos feel sterile. Real money, real trades, and a real countdown create genuine stakes that make the system's decisions meaningful.

The morning brief became the most valuable output.

The daily AI-generated email summary surfaced insights neither agent could see individually — a true meta-layer above the agents.

Human oversight is non-negotiable.

The system is autonomous but not unsupervised. Every trade is logged, every decision is auditable, and the human operator retains full override capability.

You can build for the failure you already saw and still get caught by the next one.

An earlier version of this system had no logic for a genuinely bearish regime, so it was rewritten to recognize one and stand aside. Then a market crash hit that neither agent — nor I — had anticipated. I froze trading rather than let two strategies designed for normal volatility keep operating into conditions they weren't built for. That call, not just the code, is the lesson.

The trading logic was the easy part.

The real work lived in the plumbing — order precision, fee drift, and reconciling the ledger to what Coinbase actually held.

A long-only bot can't win a bear market.

In a downturn, the winning move is to preserve capital, not to trade harder. Round 1 made that lesson unavoidable.

One shared account was the root of every headache.

Every double-sell and phantom balance traced back to two bots drawing from the same wallet. Round 2 gives every challenger their own account.