Glassbench: open-source workbench compares LLM trading agents across frameworks
davidarias2 · reddit · 2026-09-30
A developer open-sourced Glassbench (Apache 2.0), a workbench that plugs unmodified agent frameworks (TradingAgents, AI Hedge Fund) into one interface to watch, record and compare how LLM trading agents reason. Features: live per-call timelines, a searchable runs database with costs and ratings, same-stock/same-date comparisons across frameworks and LLM providers including local Ollama models, backtests vs buy-and-hold and a placebo, and Interactive Brokers paper-account order placement. 77 pre-paid runs ship with the repo so everything is readable without an API key; research use only, no real money traded.
More from coding & agent
- Cloudflare Sandboxes scale to 100k+ starts/s with sub-1s cold starts for AI agents — irvinebroque · 2026-09-30
- Google Cloud details graph workflows in ADK: fan-out, routers, and human-in-the-loop — rseroter · 2026-09-30
- Coding is still worth learning, but the focus is shifting to agent systems — lmoroney · 2026-09-30
- Dev builds night-mode coding agent that keeps working after you shut down your PC — No_Difficulty_9145 · 2026-09-30
- A custom AI agent in one line of PowerShell: your cmdlets become the agent's tools — dfinke · 2026-09-30
- User wakes up to find AI agent autonomously ideated and built VulcanBench Decision Lab overnight — gabrielchua · 2026-09-30