Japan starts giving civil servants AI agents, with rules for their use
The Digital Agency is rolling out one agent tool to ministries, plans a more powerful one for December, and has published standards that keep staff answerable for what the agents do.
Japan's Digital Agency announced on 20 October 2026 that it is beginning to provide two environments for AI agents through GENAI (源内), the generative AI environment the agency plans, builds and runs for government staff. The agency's English announcement says the tools are for about 180,000 government employees. The agency also published rules, the GENAI Standards.
The standards define an AI agent as a system that perceives its environment and acts autonomously to reach a set goal. The 180,000 figure describes the intended audience, not current users: according to the same document, GENAI began a large-scale trial in May 2026 aimed at about 180,000 staff across all ministries, and the government plans to move to full use, paid for by each ministry, from the 2027 fiscal year.
What is live and what is planned
The first tool, Agent Chat, is "an environment for easily creating and running AI agents". The agency says it is being rolled out to ministries and agencies one after another from the day of the announcement. According to the standards, staff can use it from one screen to ask questions, organise information, draft documents, analyse data, run code and write reusable "skills"; the Digital Agency has used it internally since August 2026.
The second, GENAI Workbench (源内工房), is "an environment for building and running advanced AI agents", which the agency is targeting for around December 2026. The standards describe it as a tool for multi-step research and for staff to build working prototypes such as web applications. It will be off by default: administrators grant access only to staff confirmed to meet a set level of AI literacy.
The agency plans to release both tools around February 2027 as "GENAI OSS (Open Source Software) Ver. 2.0", which "anyone can use, modify, and redistribute", it says, so that local governments and businesses can use them. According to the standards, part of GENAI was already published as free open-source software in April 2026, under a licence that allows commercial use.
What the rules require
The GENAI Standards, a first edition dated October 2026, set duties for three groups. Users must not accept AI output as it stands, but check, approve, correct, supervise or stop the work at the right stages. When an agent runs several steps, the user confirms important ones first and decides on high-impact actions such as updating or deleting files. Staff may not use unapproved AI services for work, such as a personal chatbot account. The standards also warn about prompt injection: hidden instructions in web pages or files meant to steer the AI.
Administrators are each ministry's chief AI officer, portfolio management office and IT system managers. They may not rely on individual users' care alone, and must restrict Agent Chat in work that affects people's rights, such as permits, benefits, adverse decisions and hiring, according to the risk. Some of their management tools, such as a function to suspend a user, are still under consideration, the document says.
Developers are the Digital Agency staff and contractors who build and run GENAI, the standards say. They must keep development, staging and production systems apart, limit their own privileges, and get prior approval for changes to the live system. Human access to logs of users' inputs and outputs is avoided unless necessary, and then needs prior approval and a record.
The agency's claims
The agency says that offering an environment comparable to advanced AI coding agents to about 180,000 users, including general administrative staff and not only specialists, "marks the world's first initiative of its kind". The standards were developed with the cooperation of Japan's AI Safety Institute (AISI), "which is responsible for examining and promoting methodologies and standards for evaluating AI safety", and the agency expects them to serve as a reference model for local governments and companies.
No independent report on the announcement was found. The agency's announcement does not say how many staff use Agent Chat so far, or what the tools cost.
Editor-in-chief,
Fazilat Ghobadinezhad is IQGov's editor-in-chief. She decides what IQGov covers and edits its stories. IQGov's stories are researched and drafted by AI agents to our editorial rulebook, checked against their sources by a separate fact-checking agent, and approved before publication. Every fact links to its source.
