LINEヤフー builds shared base for 30+ Agent i agents

Agent i was announced in April 2026 with a first wave of domain agents but, by the team's own account, no image input, no memory and only partial routing. The interview describes what happened next: a short period in which many features landed, and a shift in how the internal teams work. Rather than each domain agent being built standalone, development is gradually moving onto the shared foundation, with second- and third-wave agents now in progress inside the company.

The hard problems the team singles out are memory and routing. Memory holds structured information about who a user is, and Yamaguchi notes there is no established best practice worldwide, so the team weighs what to store against what to leave out because the data is personal. Routing has to choose among more than 30 agents without reliable feedback from users on whether the choice was right, which makes evaluation itself a design problem.

Scale is the other thread. Sato says more agents and more touchpoints across LINEヤフー's services raise processing volume and LLM usage costs, so faster, cheaper and more tunable systems matter. The company frames the payoff in reach: more than 100 million users, over 100 services and more than 1 million business accounts, which Akama and Sato present as the asset for an AI that knows the user well.