It is not recommended to use Agents to dance with the market because LLM models have a lot of randomness; the essence of an LLM is not a strategy engine but a probabilistic generator.

The same prompt, different outputs → unstable decisions.
You ask it the same question multiple times, and there are many answers, plus cron scripts have constraints. LLM models are stochastic, and often perform unpredictably. The worst part is memory—LLM models often forget.
Limited context window → loss of long-term state.
Memory relies on external sources (vector databases / logs) → not strongly consistent.
Retrieval itself is also probabilistic → can miss or be wrong.
And the biggest problem is that active market making (MM) is human-controlled; it has many different scripts. Most Agents can only learn from history and cannot predict the future. So using an Agent to trade volatile coins will likely end very badly.
👉 The market is not a stationary environment.
MM can switch scripts.
It can reverse your strategy to harvest.
It can create “fake historical patterns” and then cause sharp rises or crashes within your perceived regularities.
And most Agents are doing:
👉 Using historical data to fit a “stable pattern” (AutoResearch tends to overfit).
But in reality:
👉 The pattern itself is constantly being broken by people.