TraceArena
Open-source AI World OS for real-world problems
8 followers
Open-source AI World OS for real-world problems
8 followers
TraceArena is an open-source AI World OS for simulating real-world problems. Define goals, resources, rules and tools, load a scenario pack, configure your LLM, and let multiple agents compete inside one world. Every observation, decision, action, consequence and settlement is visible and replayable. Build worlds for capital markets, city governance, drug discovery, logistics or robotics.
Loaded a capital markets scenario and watched two agents actually settle trades based on the rules I set, the replay view made it super clear how decisions cascaded. Wish the docs covered more about tweaking the agent prompts though.
Thanks for trying it, Salim — glad the replay made the decision chain clear. Better prompt-tuning docs are now high on our list. We’re planning examples that separate the agent role prompt, decision policy, and world/scenario constraints. Which part felt hardest to adjust in your run? That will help us prioritize the first example.
honestly this looks really cool, the replay piece is super useful. one thing that would help a lot is letting users pause a simulation and fork it from any specific decision point so you can test how a different agent choice changes outcomes. kind of like git branches but for agent trajectories.
Love the “Git branching for agent trajectories” framing — that’s exactly the kind of counterfactual workflow TraceArena should support. We’re adding this to the roadmap: branch from a replay step, preserve provenance, and compare outcomes side by side. Would you most want to keep the world state fixed and swap one agent decision, or also edit rules/resources before rerunning?
ran a quick capital markets scenario with two competing agents and was surprised how readable the replay timeline is, you can actually follow why each side made a move. Wish the scenario pack library had a couple more ready-made worlds to poke at.
Thanks, Yaren — great to hear the timeline made the competing agents easy to follow. A larger ready-made scenario-pack library is one of the clearest requests from this launch. Which world would you explore next: logistics, city governance, drug discovery, or robotics? Your pick will help us prioritize.
Loaded a small capital markets scenario and watched two agents negotiate over a trade, the replay view made it super easy to spot where one of them misread the order book. Honestly did not expect an open-source tool to feel this polished.
Thanks, Hamdi — catching the order-book misread is exactly why we built inspectable replay: hidden agent failures become concrete evidence you can debug. Really appreciate you testing the open-source build. If any replay step was still confusing or missing a diagnostic, I’d love to hear it.