Tuned Hermes: An Off-the-Shelf Agentic Core Shaped to ECOM1

Editorial Note from Rinat

Key lesson of this architecture: don’t build the entire agent from scratch if you can tune a strong existing agentic core.

Ivan took Hermes, narrowed the tool surface, shaped it with a precise prompt and a few skills, and routed only the harder task cluster to a stronger model. The linked competitive run scored 0.5735 with 53 perfect trials, while later Hub runs reached 0.8105.

BitGN platform logs all agent interactions, so we can see where the architecture was strong and where it still leaked points. The main gotcha (and an opportunity to improve quality further) was citation discipline. In the linked run, most non-perfect trials had reference/citation failures, usually extra refs rather than missing ones, especially under /proc/catalog and /proc/carts.

Why do the citations matter? It is the one lesson that I’ve learned from shipping LLM-driven systems in the enterprise environments: presence of correct grounding references correlates with the correctness of the final answer from the agent (and references are can be verified easily in test harnesses). The lesson is so important, that my video course on AI foundations invests extra time to dive deeper into the topic.

The read vs. read_silent idea was directionally right, but brittle: one wrong tracked read could sink an otherwise correct answer.

However these nuances are not that important, given the fact that this architecture was built in “hardcode mode” - using LLMs that could be run locally:

• DeepSeek V4 Flash - Mixture-of-Experts with model 284B parameters (13B activated) - HF
• DeepSeek V4 Pro - Mixture-of-Experts model with 1.6T parameters (49B activated) - HF

Both models support context length of 1 million tokens, and are pre-trained on 32T tokens.

AI Engineers in the circles especially praise V4 Flash for good efficiency-per-dollar ratio.

I think, this architecture will make a good starting point for ECOM2 as well, with a few possible improvements:

• keep the tuned-Hermes shape - it works well;
• add deterministic pre-submit reference validation by task family (don’t overfit though, just route and reason, potentially with Schema-Guided Reasoning);
• catalogue, carts, payments, staff, stores, and exports each need different minimal evidence, pay attention to that;
• pair that with small verifier code for export equality, dispatch JSON validity, and count/resolve/fraud tasks, and this architecture becomes much more production-like without losing its useful simplicity.

Congratulations to Ivan Konovalov for showing that a stock agentic core plus disciplined tuning can go far on open-weight models. Here is his GitHub repository with the source code.

Read the architecture writeup, crafted by the team behind this tuned Hermes agent: Ivan Konovalov and his Claude Code agent!