Zafin signs definitive agreement to acquire Naehas

x

The Ten Commandments for Governed Agentic AI 

Ten architectural principles for controlling autonomy, preserving accountability, and building trust in enterprise AI systems. 

The move from generative AI to agentic AI fundamentally changes the nature of enterprise risk. The question is no longer simply what an AI model said, but what the system did, what authority it had to act, and whether the organization can account for its actions. 

AI agents can retrieve information, invoke tools, coordinate with other agents, and execute multi-step workflows. They can modify customer records, change product configurations, execute code, or initiate financial transactions. These capabilities promise significant operational benefits, but introduce risks that traditional model governance was not designed to address.

Adoption is accelerating. According to the Bank of England and FCA’s 2026 survey, 80% of responding financial firms have deployed AI agents, although only 2% report giving agents broad autonomy. 

Meanwhile, regulatory expectations are evolving. Singapore’s Model AI Governance Framework for Agentic AI explicitly addresses agentic systems, while the Monetary Authority of Singapore’s October 2026 guidelines establish AI risk management expectations for financial institutions. Canada’s OSFI Guideline E-23, effective May 2027, extends model risk management requirements to AI and machine learning. The EU AI Act introduces obligations around risk management, logging, and human oversight for qualifying high-risk systems, with key requirements applying from December 2027. 

These frameworks differ in scope, but point toward a common challenge: institutions must maintain control over AI systems that increasingly operate across applications, data sources, and organizational boundaries. 

Meeting that challenge requires governance of the entire execution environment, not just the underlying models. 

The following ten commandments establish three foundational architectural disciplines for governed agentic AI.

Theme 1: Define and control the scope of what is governed 

The first challenge is controlling what agents can access and do. 

In April 2026, a coding agent at PocketOS reportedly deleted a production database and its backups in approximately nine seconds. The incident exposed a fundamental weakness: an agent working on a legitimate task had access to permissions that enabled destructive actions far beyond its intended scope. 

Such failures demonstrate why instructions alone are insufficient. Governance must be enforced through architecture, identity, authorization, and controlled access to enterprise systems. 

1. Govern work, not just agents 

The fundamental unit of governance should be the business task, rather than the individual agent executing it. 

Consider a workflow to adjust a customer fee. Multiple agents might retrieve customer information, calculate an adjustment, check eligibility, and prepare an update. Human approval may also be required. 

Policy, authorization, and evidence must remain attached to that workflow regardless of which participant performs each step. 

This requires persistent workflow identities, defined execution boundaries, policy enforcement points, and orchestration that maintains context across agents and systems. Governance should follow the work, even as the agents performing it change. 

2. Know every agent by name, owner, and authority 

An organization cannot govern agents it cannot identify. 

Every agent should have a distinct identity, documented purpose, accountable owner, and explicitly defined permissions. This applies equally to internally developed agents, third-party applications, and agents operating across organizational boundaries.

A governed architecture requires an enterprise agent registry recording ownership, capabilities, dependencies, risk classification, and authorization scope. 

At runtime, agents should authenticate through non-human identities and receive narrowly scoped credentials. When one agent delegates work to another, permissions must be evaluated independently rather than inherited without restriction. 

Discovery and inventory establish visibility. Identity and authorization controls establish accountability. 

3. Match autonomy to risk 

Autonomy should be an explicit policy decision, not an assumed property of an AI model. 

An agent summarizing internal documentation presents different risks from one modifying customer entitlements or initiating financial transactions. 

Architectures should support graduated autonomy, from recommendation-only operation through constrained execution to conditional or greater independence for appropriately bounded tasks. 

Permissions must reflect potential impact, reversibility, data sensitivity, and operational context. Controls may include transaction thresholds, restricted tool access, execution limits, and mandatory approvals. 

Autonomy should also be reassessed as agents demonstrate reliability or encounter new risks. 

The objective is to grant sufficient autonomy to create value without surrendering control over consequential actions. 

4. Control tools, protocols, and external systems 

Tools are where agent intentions become operational consequences. 

Every external API, integration, and Model Context Protocol (MCP) server should be treated as a governed asset, with an owner, approved purpose, permission scope, and security assessment. 

MCP standardizes connections between AI applications and external resources, but does not inherently establish the trustworthiness of those resources or the safety of their outputs. 

Institutions need enforceable controls at these boundaries, including allowlisted operations, scoped credentials, input validation, and restrictions on external connections. 

Untrusted tool responses must also remain separate from authoritative instructions to prevent prompt injection and unauthorized execution. 

Agents should never acquire authority simply because they can discover or connect to a tool. 

Theme 2: Keep humans in command of the work 

As agents execute increasingly complex workflows, human accountability becomes harder to maintain through conventional supervision. 

Legal precedent reinforces the importance of this responsibility. In 2024, a Canadian tribunal held Air Canada responsible for misleading information supplied through its chatbot. Agentic systems extend the challenge beyond information into consequential actions. 

Singapore’s governance framework emphasizes meaningful human accountability, while emerging regulatory requirements increasingly focus on oversight, traceability, and operational controls. 

The architectural challenge is to make human authority enforceable without requiring manual supervision of every individual action. 

5. Preserve human authority where consequences matter 

Human oversight must be meaningful, proportionate, and embedded within execution. 

High-impact, irreversible, or exceptional actions should require explicit authorization from an appropriately empowered individual. 

Approval mechanisms must provide sufficient context to make an informed decision: the proposed action, supporting evidence, potential consequences, and available alternatives. 

An approval checkpoint should suspend execution until authorization is received. It must not be bypassable through alternative tools or agent-to-agent delegation. 

Institutions must also account for automation bias and alert fatigue. Human involvement is effective only when reviewers understand what they are approving and retain genuine authority to intervene. 

6. Create proof as work happens 

Auditability cannot depend on reconstructing an agent’s activities after an incident. 

Evidence should be generated continuously as part of workflow execution, capturing the identity of participating agents, authorized objectives, relevant inputs, tool invocations, policy decisions, approvals, and resulting changes. 

This requires structured, tamper-resistant execution records that connect actions to the business process and authority under which they occurred. 

Logs should preserve sufficient context to reconstruct consequential decisions without unnecessarily exposing sensitive information. 

The objective is more than observability. Institutions must be able to demonstrate that an action was authorized, executed within defined boundaries, and completed as intended. 

7. Monitor, intervene, and fail safely 

Pre-deployment controls cannot anticipate every possible agent behavior. 

Runtime governance must therefore detect policy violations, anomalous tool use, unexpected execution paths, and other indicators of compromised or unreliable operation. 

Monitoring should operate at both the individual agent and workflow levels, since apparently acceptable actions may create unacceptable outcomes when combined. 

Institutions need mechanisms to pause workflows, revoke credentials, isolate agents, and prevent further actions when defined thresholds are breached. 

Where possible, architectures should support rollback or compensating transactions. 

The governing assumption should be that failures will occur. The system must detect them quickly, contain their impact, and preserve evidence for investigation. 

Theme 3: Earn and sustain trust 

Agent reliability is not established permanently at deployment. 

Large language models are probabilistic, and agent behavior depends on changing combinations of prompts, models, tools, data, memory, and external services. 

The Bank of England’s 2026 research identifies oversight of third-party AI systems, ongoing monitoring, and change control as significant governance challenges. 

Trust therefore requires continuous validation across the complete agent lifecycle. 

8. Test before trusting 

Agents must be evaluated against the conditions in which they will actually operate, including conditions they may encounter unintentionally or through malicious interference. 

Testing should examine more than response accuracy. It must assess tool execution, authorization boundaries, data access, policy compliance, and behavior across multi-step workflows. 

Adversarial testing should include prompt injection, malicious tool outputs, unexpected inputs, attempted privilege escalation, and failures involving multiple agents. 

Higher-risk deployments demand stronger evidence of safe behavior before authorization. 

Evaluation should also continue after deployment, using representative operational scenarios and clearly defined acceptance thresholds. 

Trust must be demonstrated through testing rather than inferred from model performance alone. 

9. Govern change, version, and drift 

An agent is a configuration of interconnected components, not simply a model. 

Changes to the model, system prompts, tools, permissions, memory, knowledge sources, or execution policies may alter its behavior. 

Governed architectures need version control across these dependencies, documented

change approval, regression testing, and mechanisms to detect behavioral drift. 

Material changes should trigger reassessment of the agent’s permissions and risk classification before deployment. 

Third-party updates require particular attention because changes may occur outside the institution’s direct control. 

Effective lifecycle governance must make changes traceable, testable, and reversible where technically feasible. 

10. Prevent AI sprawl from becoming operating-model sprawl 

When teams deploy agents independently, institutions risk creating fragmented operating environments with inconsistent controls, duplicated workflows, and limited enterprise visibility. 

The challenge extends beyond identifying every agent. Organizations must understand how agents interact, which processes they support, and where responsibility resides. 

A scalable architecture should provide shared capabilities for identity, policy enforcement, orchestration, monitoring, and evidence collection across different agent frameworks and providers. 

Governance need not require every agent to operate on the same platform. It does require consistent standards and clearly defined integration boundaries. 

Without these foundations, agent proliferation can become an ungoverned parallel operating model. 

From autonomous agents to governed enterprise systems 

Agentic AI shifts the governance challenge from controlling individual model outputs to controlling actions across interconnected systems. 

Incidents involving destructive operations, unauthorized access, and AI-enabled cyberattacks demonstrate that traditional model-level safeguards are insufficient when software can act independently. 

Regulatory frameworks are beginning to address these risks, but institutions cannot depend on regulation alone to define the necessary architecture. 

The ten commandments provide a foundation: govern work and permissions, preserve meaningful human authority, and continuously validate the systems executing enterprise processes. 

Ultimately, the goal is not to eliminate autonomy. It is to establish the architectural controls that make autonomy trustworthy, accountable, and sustainable at enterprise scale. 

Sources and references 

The following research, regulatory guidance, industry frameworks, and reported incidents informed the architectural principles presented in this article. 

Regulatory and governance frameworks 

1. Infocomm Media Development Authority (IMDA), Singapore 

Model AI Governance Framework for Agentic AI, Version 1.5 (May 2026) 

Read the framework 

Provides agent-specific guidance on bounding risks, meaningful human accountability, technical controls, and responsible deployment. 

2. Monetary Authority of Singapore (MAS) 

Guidelines on Artificial Intelligence Risk Management for Financial Institutions (October 2026) 

Read the guidelines 

Establishes supervisory expectations for financial institutions, including AI governance, risk assessments, lifecycle controls, and third-party oversight. 

3. Office of the Superintendent of Financial Institutions (OSFI), Canada 

Guideline E-23: Model Risk Management (September 2025; effective May 2027) 

Read the guideline 

Sets expectations for model identification, risk classification, validation, monitoring, and lifecycle governance, including AI and machine learning models. 

4. European Commission 

EU Artificial Intelligence Act: Implementation timeline (updated 2026) 

View the implementation timeline 

Outlines phased regulatory obligations, including risk management, human oversight, transparency, and logging requirements for applicable AI systems. 

Industry research and technical standards 

5. Bank of England and Financial Conduct Authority (FCA) 

Artificial Intelligence in UK Financial Services – 2026 (October 2026) 

Read the report 

Documents AI agent adoption, degrees of autonomy, governance practices, third-party dependencies, and emerging operational risks across UK financial services. 

6. OWASP GenAI Security Project 

OWASP Top 10 for Agentic Applications for 2026 (December 2025) 

Read the framework 

Identifies critical security threats affecting autonomous AI systems, including tool misuse, agent identity and privilege abuse, and agent goal hijacking. 

7. Model Context Protocol 

Model Context Protocol Specification (2026) 

Read the specification 

Defines the technical protocol for connecting AI applications with external tools, data, and services, including relevant security and authorization considerations. 

Reported incidents and legal precedent 

8. TechSpot 

AI Coding Agent Running Claude Wiped a Startup’s Database (and Its Backups) in 9 Seconds (April 2026) 

Read the article 

Reports the PocketOS incident, illustrating the consequences of excessive agent permissions, inadequate environmental isolation, and uncontrolled tool execution. 

9. American Bar Association 

BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot (February 2024) 

Read the analysis 

Examines the Moffatt v. Air Canada decision and its implications for organizational accountability when AI systems provide incorrect information. 

10. Anthropic 

Disrupting the First Reported AI-Orchestrated Cyber Espionage Campaign (November 2025) 

Read the report