Digital Agency Opens AI Agent Tools to 180,000 Government Staff in Gennai and Publishes Safety Rules
Japan’s Digital Agency has started giving central government staff the ability to build and run their own AI agents on Gennai, the in-house generative AI environment used by about 180,000 officials, Impress Watch reported on October 21. An agent chat tool began rolling out to ministries and agencies on October 20, and a more powerful development workspace called Gennai Kobo is due by around December.
The agent chat is pitched as the next step beyond a conversational assistant: staff describe a task in plain language, and the agent strings several steps together, consults files, searches the web and returns results as documents. The agency compares it to Claude Cowork. Gennai Kobo, which it likens to Claude Code, gives each user a personal web development environment that can visit any site, run commands, handle files and chain dozens to hundreds of steps, enough to build working prototypes. A third element, Gennai Library, will let staff share reusable skills across ministries. The Digital Agency says putting tools of this kind in front of ordinary civil servants at this scale is a world first.
Alongside the rollout it published the Gennai Standard, a set of governance, security and safety rules drafted with help from the AI Safety Institute. The rules split duties among users, administrators such as each ministry’s chief AI officer, and the agency’s own developers. Officials must understand an output’s basis and limits before relying on it, back up important files, narrow an agent’s scope, verify the results, and stop and report anything abnormal. They should not run processes they do not understand.
What it means in practice
The move turns the government into one of Japan’s largest test beds for workplace agents, a shift already visible in companies such as NTT Docomo Business and its AI bosses. It also gives practical weight to the Prime Minister’s push to spread high-performance AI, while drawing on the AI Safety Institute’s work on domestic model benchmarks.
Risks are spelled out rather than played down. The standard warns that an error early in an agent’s chain can cascade into later steps, that unchecked outputs used in official notices could harm citizens’ rights and interests, and that direct web access combined with long autonomous runs creates a route for supply-chain contamination and unintended data leaks.
What to watch: how many ministries switch on the agent chat in the coming weeks, whether Gennai Kobo arrives on schedule, and the planned release of both tools to local governments and companies as open-source Gennai OSS Version 2.0 around February 2027.