Problem
Yandex Lavka had many routine processes done by hand across 6+ admin panels, a tracker and chats. One was investigating lost parcels: find duplicate tickets, fill a 30+ field form, collect missing data, and issue a resolution by a 5-page instruction (was it lost, and whose fault). The process took 13 full-time equivalents (FTE); at 70% automation the expected gain was 9 FTE. A naive ReAct agent with tools fails on legacy admin APIs, a contradictory instruction and data that arrives hours late.
Example from the article: at the first launch the primary form was fully correct in 10% of cases and almost empty in 26%. Resolutions were correct in 11.5%.
Solution in brief
Rank AI projects by FTE gain per cost and run all agents on one shared platform, where each agent needs only a prompt, logic and an eval set. For the hard process, split one agent into three, rewrite ReAct as a StateGraph and move everything deterministic into code, leaving the LLM only what code cannot do.
Questions this case answers
- How to turn a manual investigation that spans several legacy admin systems, a long form and a written policy into a reliable LLM agent
- How to decide between a free-running tool-calling agent, a multi-agent system and a deterministic graph for a process that is hard even for people
- How to rank candidate automation projects by labor saved per unit of cost and measure the real effect after launch
- How to turn a manual support investigation (ticket, 6+ admin panels, 30-field form, verdict) into a production LLM agent with high accuracy
Tools and methods
- FTE gain — project value in "virtual employees" (160 hours per month each)
- ReAct agent → StateGraph — free tool choice replaced by a graph with fixed nodes and transitions
- LangGraph runtime — runs the graph with checkpoints
- Agent platform — shared interfaces, 118 tools, auth, LLM routing, memory, queues, eval
- Eval basket — per-agent set of cases for quality tracking in production
Google or a one-time email code. No password.
