Building complex research systems often starts with massive ambition, resulting in staggering hardware bills and architectural headaches. Prokhor, lead of the "Explore" agent team (the deep research mode in Alice AI), shared practical insights on deploying a heavy agent pipeline into production. On the product side was product manager Ruslan Iliev: he outlined goals, assembled a toolset, and prepared a validation dataset, driving the launch from a closed waitlist to production. Development started in the spring when the market was hit by a wave of reasoning models, prompting the team to roll out the "Reason" mode.
Abandoning training from scratch
Agent scenarios inherently burn through resources: the system makes hundreds of search queries, processes dynamic JS content, and runs Python code for calculations. In the early iterations, response generation times reached half an hour. To avoid going broke on infrastructure, the engineers consciously abandoned fine-tuning large base models from scratch.
"I used to train models from scratch for information extraction and semantic search. Then GPT-3 and FLAN-T5 appeared — and my models became obsolete overnight."
This eternal technological dilemma is phrased the exact same way across all mature teams. Instead of rewriting weights for every single task, developers bet on orchestration, context engineering, and auxiliary models.
Architecture shift and results
The first prototype was built following a CodeAgent pattern, where the model invokes tools by generating Python code. Benchmark numbers looked decent: 75 on FRAMES, 81 on SimpleQA, and 26 on GAIA (the low score was explained by the agent's initial inability to handle files). However, security blocked this architecture from reaching production. Executing code in an isolated sandbox with internet access required proxying calls outward, which looks like a minefield for information security.
The situation changed after the release of open-source models with proper function calling. In just two weeks, engineers rewrote the agent using a classic function-calling loop. The previous CodeAgent metrics were surpassed within a week, the infrastructure became several times simpler, and security concerns stopped giving the team nightmares.