The Cited Sandbox: One Open-Weight Model, One Tool, Eight Gates

Editorial Note from Rinat

Let’s explore an experiment on architecture of a solution for Agentic Commerce from Farid Temuri.

This agent scored 71.8 ECOM1 Ultimate leaderboard (during the 3-hour blind run) and got in TOP-20. The most impressive part is - it achieved that with a simple architecture and xiaomi/mimo-v2.5-pro open-weight model under the hood.

The architecture involved these key blocks:

• LLM was running in a single agentic loop (REPL) with the feedback
• LLM had access to the sandbox, where it could run code, read files, access scratch pad.
• Results had to go through the submission gates with a set of deterministic checks
• Agentic responses that failed the submission gate - would be handed back to the REPL loop with more context about the needed repair

So all together: one tool, no planner, no router, no LLM judge, no fine-tuning. The agent was allowed to write real code against the runtime instead of calling narrow tools. The harness also guarded agent against submitting an answer that isn’t grounded in files it actually read.

Thanks to this architecture, agent was able to avoid many no-answer and hallucinated-reference failures that are common among agents attempting to solve ECOM1 tasks.

BitGN platform analysis shows that the architecture still had a few weak points: occasional missing or extra required refs, wrong cited SKU families, troubles with OCR uploads, carts, and checkout-policy files. The agent studied the context before answering, but the evidence contract was not always aligned with what the business process expects.

The model under the hood is - xiaomi/mimo-v2.5-pro:

1.02T-parameter Mixture-of-Experts model with 42B active parameters, built on a hybrid-attention architecture with a 1M-token context window.

The model is larger than usual, yet it can still run locally on a modest GPU cluster, thanks to the MoE architecture.

This agent would be a good baseline for the next ECOM2 round. It has:

• filesystem-first navigation
• task-specific citation checklists
• strict outcome-class rules
• exact answer-format gates

If those are added without bloating the stack, this style of agent could become a strong open-weight reference point for ECOM1-style work.

Congratulations to Farid Temuri for scoring top-20 on the blind bitgn/ecom1-prod leaderboard (run-22RxPyYQ4dtnsaeKdXpRsJ6ce) using one open-weight model (xiaomi/mimo-v2.5-pro) and a very straightforward architecture. Here is his Github repository with the source code (TypeScript / Bun).

Read the agent summary, written by the team that created this agent: Farid and his Claude Agent.