BetterTradingAgents swaps CrewAI for OpenAI's SDK and fixes paper sells
Version 0.3.0 changes the research-agent harness, caps model-call concurrency and switches paper orders from dollar notional to shares.
BetterTradingAgents 0.3.0 replaces CrewAI with the OpenAI Agents SDK and fixes an accounting detail in its paper-trading workflow. Published October 27, the release changes how the standalone research application coordinates agents and how simulated orders reach its broker integration.
The SDK migration does not make OpenAI the only model provider. The Python application documents OpenAI-compatible endpoints including DeepSeek, OpenRouter, Z.AI, vLLM and Ollama, with role-specific overrides. Six research roles examine market regime, technicals, fundamentals, news, sentiment and forecasts; deeper runs add bull and bear arguments and a neutral judge before a risk gate.
The maintainer reports about 32% lower cost per run in a live comparison with the release build, with none of six decisions changing. That is a small author-run comparison, not a broad benchmark or evidence of profitable trading. A new configurable concurrency cap, defaulting to 12, also limits request fan-out to reduce rate-limit errors.
The paper-order correction is concrete: orders now use share quantities instead of dollar notional. The release says price conversion could previously oversell a full-position exit; the server now clamps sells to the shares held. The review screen adds a six-month chart and compares live prices with decision-time prices.
To try the local app, the pinned README requires Python 3.12 or newer and uv, followed by uv sync, uv run setup and uv run app. It binds to localhost on port 8000. Users supply their model credentials and inference costs; without an LLM API key it enters a labeled mock mode using live market data. Optional Alpaca credentials enable paper trading behind a default-off switch. The documented client uses paper mode and has no real-trading path; simulated decisions do not establish investment returns.