On 20 October 2026, Japan's Digital Agency published the 'GENAI Standards', a set of governance, security and safety requirements for AI agents used inside government. According to the Digital Agency's GENAI page, the release came with two new tools on the government AI platform: 'Agent Chat' and 'GENAI Workbench'. The agency says it expects local governments, companies and other organisations to use the standards as a reference model. An open-source release of the software, GENAI OSS Ver. 2.0, is planned for around February 2027, according to ActuIA's report.
The most interesting feature is not any single rule. It is the order of operations: the rulebook arrived alongside the tools, not years after them.
What the standards actually do
The standards page sets out 'fundamental principles and compliance requirements for relevant parties' and splits them across three roles: Users, Administrators and Developers. The agency's stated reasoning is that AI agents can do more than conventional AI applications, so they need clear rules to stay secure and well governed. The standards were developed with the cooperation of Japan's AI Safety Institute (AISI), which examines and promotes methods and standards for evaluating AI safety.
The agency's page does not publish a user count or detail specific technical controls, and this article does not speculate on either. What can be said is structural. Responsibility is assigned by role, rather than by a single abstract duty on 'the system'. The person who uses an agent, the person who administers the environment and the person who builds the software each face their own obligations.
The case for stricter rules
The strongest argument for heavier regulation of agents is that they differ in kind from chatbots. An assistant that only drafts text leaves a human to act on it. An agent that can take actions on its own can compound a mistake across systems before anyone notices. Governments hold sensitive data and make decisions that affect citizens' benefits and rights, so a cautious, prescriptive approach is defensible, and some regulators elsewhere have urged restraint on agentic products in production environments. Critics of soft-law approaches also note that voluntary guidance has no enforcement bite.
That argument has real force when the deployer is the state, because citizens cannot opt out of a government system.
Why this model is the proportionate answer
But Japan's choice is instructive precisely because the government is regulating itself first. The GENAI platform has been built in-house since its launch in May 2025, initially for the Digital Agency's own employees, with rollout to other ministries starting in January 2026 on a trial basis and full-scale use scheduled from fiscal year 2027, according to the agency's GENAI page. That is a staged path, where the rules are tested on a bounded population before anyone proposes extending them to the private sector.
This matters for a pro-innovation reading of tech policy. The common failure of AI regulation is writing binding obligations for a technology whose behaviour is still being learned. Rules drafted against an imagined risk tend to either miss the real failure modes or freeze one architecture into law. Standards that govern a live deployment, written with a safety institute that evaluates real systems, can be revised as the evidence comes in. The fiscal 2026 pilot year is effectively a feedback loop.
The approach also fits Japan's broader legal posture. Japan's AI Promotion Act, approved by the Diet on 28 May 2025, is an innovation-first framework. As the Future of Privacy Forum's analysis notes, it 'contains no explicit penalties, financial or otherwise, for non-compliance', asks businesses only to 'endeavor to cooperate' with government measures, and relies on guidance and voluntary initiatives. The GENAI Standards are consistent with that design: they give the public sector a concrete benchmark without creating a new compliance regime for private firms.
Reference model, not mandate
The key phrase in the Digital Agency's framing is 'reference model'. The agency expects local governments, companies and other organisations to consult the standards, not to be bound by them. That distinction deserves defence. A mandate would harden early assumptions across thousands of organisations with very different risk profiles: a municipal office and a manufacturer should not necessarily run the same agent controls. A reference model lets each adapt the framework to its own context while benefiting from tested vocabulary, such as the User, Administrator and Developer split.
There is also a competitive benefit. If the open-source release arrives as planned around February 2027, organisations will be able to inspect and adapt the same software the government uses, which lowers the cost of building responsibly. Open code plus an open rulebook is a better way to spread good practice than a licensing regime.
What to watch
There are honest open questions. Voluntary reference models only work if people actually adopt them, and the agency has not published adoption data. The standards' real quality will only be visible once the technical detail and the fiscal 2026 pilot results are examined by outside reviewers. Transparency about incidents, and about what changes after the pilot year, would let others judge whether the standards work in practice and not merely on paper.
There is also a risk of drift: a reference model written for government can quietly become a de facto procurement requirement for vendors, which would turn a soft standard into a hard one without any democratic debate. Policymakers elsewhere should watch for that.
For now, the lesson is modest and useful. Governance of AI agents does not have to start with a statute that bans or licenses. It can start with the institution that is deploying them, assign responsibility by role, involve a safety institute, publish the result, open-source the code and iterate. Other jurisdictions weighing agent rules would do well to copy the sequencing, even if they write different substance.