The Government of Canada has presented new guidance to help federal departments and agencies with the adoption, evaluation and responsible governance of agentic Artificial Intelligence (AI) systems. The document represents one of the first institutional frameworks designed specifically to prepare the public administration for the arrival of systems capable of acting autonomously, planning complex tasks and executing actions with limited human oversight.
The guidance published by the Canadian Government serves as a transitional regulatory and operational framework intended to help public institutions take advantage of advanced automation without losing control over the ethical, legal, privacy and security risks associated with this new generation of AI. The goal is to ensure that these technologies remain aligned with the Directive on Automated Decision-Making and with the governance frameworks already in force in the country.
Traditional generative AI and agentic AI
One of the central points of the document is the distinction between traditional generative AI and so-called agentic AI. According to the guidance, conventional generative AI works reactively: it responds to specific instructions by generating text, code or images, but it needs a human to subsequently decide what to do with those outputs.
Why does AI consume 1000% more energy than a traditional Google search?
Agentic AI, by contrast, is defined by its ability to act proactively and independently to achieve complex long-term objectives. These agents can perceive their environment, plan multi-step workflows, make delegated decisions, interact with other agents and use external tools with reduced or intermittent human oversight.
To help public institutions assess the impact and degree of risk of these systems, the Canadian Government designed a taxonomy based on four progressive levels of autonomy. The example used in the document is the organization of work meetings.
Several levels
At Level 1, called Assistive, the system only identifies free slots in calendars and suggests times, while the human user remains responsible for selecting the time and sending the invitations. At Level 2, Semi-autonomous, the system already proposes optimal times and drafts invitations, although it still requires explicit human approval before executing any action.
Level 3, High Autonomy, represents a significant leap: the agent receives a general goal and automatically executes actions such as booking spaces, sending invitations and logging all its operations, notifying the human user once the process is complete. Finally, Level 4, Adaptive Autonomy, allows the system to react in real time to changes or conflicts, renegotiating and rescheduling tasks autonomously and escalating only complex exceptions to a human operator.
The Government of Canada considers that this type of AI agent can offer significant operational advantages for modernizing the public administration. Among them it highlights the massive automation of repetitive and complex tasks, the standardization of administrative processes through executable digital templates and the freeing up of time for public employees to focus on functions that require critical judgment, empathy and direct attention to citizens.
Automation bias
However, the document also warns about the risks associated with the implementation of autonomous systems. One of the main dangers identified is the possibility that small logical errors or incorrect interpretations trigger massive erroneous actions at high speed, such as improper modifications of critical records or mistaken sending of official communications before a human supervisor can intervene.
Security vulnerabilities
The guidance also warns about security vulnerabilities arising from instruction injection or prompt injection attacks. Because these agents process unstructured external data, such as emails, forms or attachments, they could be manipulated through hidden malicious instructions aimed at forcing the bypass of security restrictions, the deletion of data or the execution of hostile actions.
Another of the risks identified is so-called automation bias, which could lead public staff to accept the recommendations of AI agents uncritically. The document warns that excessive dependence on these systems could progressively weaken professional capabilities and reduce the institutional ability to detect errors or biases.
On the legal front, Canada also acknowledges that the use of autonomous agents in processes that affect citizens' rights, benefits or legal status poses significant transparency and accountability challenges. The guidance holds that any system that decisively influences an administrative decision must be able to explain clearly and accessibly the logic behind its decisions so as not to undermine the principles of procedural justice.
Bounded Autonomy
To respond to these challenges, the document introduces two new governance principles specific to agentic AI environments. The first is Bounded Autonomy, which establishes preventive limits on the operational independence of agents. This includes explicit restrictions on the databases, APIs and tools the system can use, as well as visible indicators that make known at all times the level of permissions granted to the agent.
The guidance also requires that each deployed agent have a clearly identified human responsible party, who will assume the legal and institutional responsibility for the actions carried out by the system and by any subagents automatically delegated.
The second principle introduced is Recoverability, focused on resilience and damage mitigation. The Canadian Government requires that all systems have immediate stop mechanisms or kill switches, capable of instantly pausing or revoking the credentials and operations of an autonomous agent.
Likewise, it mandates maintaining immutable audit logs where all decisions, tools used and actions executed by the AI are stored chronologically, in hardened repositories that the agent itself cannot access to modify or delete information.
Progressive adoption
The guidance also establishes fail-safe architectures for those administrative processes where an automated action may be irreversible or very difficult to correct. In such cases, the system must block direct execution and generate previews or work queues that require human validation and digital signature before any definitive action is carried out.
Finally, the Canadian Government proposes a progressive and iterative adoption of these technologies. Federal institutions must begin with comprehensive impact assessments to identify possible effects on citizens and initially deploy controlled pilots with low levels of autonomy. Only after empirically verifying the effectiveness of security and recovery controls may they move toward high-autonomy models.
The publication of this guidance places Canada among the first governments attempting to develop a specific institutional framework to regulate the arrival of autonomous AI agents in the public sector, in a global context where the transition from generative AI to systems capable of executing real actions is beginning to accelerate debates on governance, accountability and technological control.





