Agent design
Most LLM trading agents look brilliant in a demo and fall apart in practice. The research on what actually separates strong agents from weak ones is now clear, and it is mostly about discipline, memory, and honesty rather than a clever prompt. Here is what to build, and why, on CoinRithm's paper-trading proving ground.
The loop
Every good trading agent runs the same feedback loop. The quality is in how disciplined each step is, not how clever any single one looks.
Pull ground truth from tools: live price, your positions and cash, news, and market context. Read before you reason, and only read what was knowable at decision time.
Reason explicitly. State a thesis, a probability, and a risk-reward, so the decision can be checked later instead of trusted blindly.
Quote before you trade, then place a validated order. The action shape is fixed and checked, so an invalid trade can never reach the engine.
When a position closes, make the agent compare its thesis to the outcome and store the lesson. Tomorrow's decision should be able to retrieve today's mistake.
What works
These are the most-replicated findings across the agentic-trading literature and the 2025-2026 benchmarks. None of them is exotic; all of them are usually skipped.
Keep a working memory of the current situation plus a longer-term store of past trades, ranked by how recent and relevant they are.
Stratified, decayed memory is the single most-replicated win across FinMem and TradingGPT. The same model with a better memory makes different, better decisions.
On each close, force the agent to critique its own thesis against what actually happened, and write that lesson back into memory.
Reflection loops that store outcome critiques are what let an agent improve instead of repeating the same mistake. CoinRithm already captures closed-position feedback to build on.
Run a separate bull case and bear case, then a neutral judge that picks. Do not just ask one model to weigh both sides.
Forced adversarial debate (TradingAgents) beats single-model both-sides reasoning, which tends to rationalise the answer it already had.
Put position caps, a leverage ceiling, and a max-drawdown stop in code, not in the prompt. On rejection, send the agent back for a lower-risk plan.
Agents with a hard risk layer kept lower maximum drawdown even when their returns were modest. A rule in the prompt is a suggestion; a rule in code is a rule.
Let the agent hold. Make not-trading on a weak signal a valid, encouraged action.
Disciplined agents in the benchmarks traded roughly 14 to 32 times; compulsive ones spiked to 85 to 101 and destroyed their own returns. Trade count does not predict profit.
Require an explicit probability on every call, size positions by it, and track calibration over time (is a stated 70% right about 70% of the time?).
LLMs are persistently overconfident at high probabilities. A well-calibrated agent is a more honest skill signal than raw profit, which is mostly market exposure and luck.
Feed it news and context, not only price. Build the default agent on multiple sources.
Removing news and fundamentals collapsed a top benchmark agent's return from 1.9% to 0.6%. Single-source agents reliably underperform.
Only let the agent see information that existed at decision time. Be especially careful with retrieved memories that mention how an event turned out.
The 'Oracle Fallacy': agents that retrieve future-tainted context stop predicting and start remembering. Point-in-time discipline is what separates a real result from a leak.
Prediction markets
Binary markets reward a different skill than spot or futures. Accuracy is not enough; you need an edge over the price.
On CoinRithm
CoinRithm is a proving ground, so the pieces that make these principles checkable are built in.
These are research-backed design heuristics for a paper-trading sandbox, not financial advice. Past or simulated performance does not predict real-money results.