DS-Agent is a multi-agent system, built inside Bank of Georgia, that changes how the bank delivers data science. A small human-led team sets the direction and keeps the decisions, while more than ten specialised agents and micro-agents carry out much of the planning, investigation, modelling, validation and deployment. The architecture stops the system passing any significant milestone without a person approving it. The principle behind it is humans lead, agents execute.
It emerged from a capacity problem. Demand on the data analytics team ran well ahead of what it could deliver, and hiring more people would have added coordination cost along with capacity. Rather than make individual data scientists faster, the team changed the unit of delivery: a small human pod working with an agent crew.
On its first benchmarked project, work estimated at four people over twelve weeks was completed by two people in eight, a threefold gain in efficiency (once time and headcount are taken together), and the resulting models were judged better than the conventional build. Two projects have now been delivered this way, with six more in development.
The more interesting effect has been on how the team experiments. Questions that would have cost a week can now be explored in an afternoon, so approaches that would once have been dismissed as too expensive to try are worth testing. Human review is deliberately placed where judgement matters rather than everywhere, which keeps people accountable for the decisions without putting them back in the middle of every step.
