Agentic AI Governance Frameworks 2026: Emerging Standards, Risks and Insights (part 2)

Article published on Hackernoon

It was clear that Agentic AI would dominate the next phase of the AI discourse. Early harbingers of this emerged in January, with the release of the first Agentic AI framework by Singapore’s Infocomm Media Development Authority (“IMDA”).

In February, I published an article reviewing the key Agentic AI governance frameworks that emerged in the previous couple of months, ranging from governance-focused standards to highly technical programmes.

Less than three months later, the number of frameworks has rapidly multiplied. Beyond confirming the growing relevance of AI agents as core business accessories, this also signals strong demand from companies for practical implementation models. Based on the number and the nature of frameworks surfaced since February, it appears that Agentic AI governance has shifted from a vague and somewhat self-referential debate to system-level control architectures.

In this article, I’ll summarise the key developments in Agentic AI over the past three months, focusing on expert insights, risk and governance considerations, and emerging standards.

Consultancies leading the Agentic way

Professional services firms demonstrate strong awareness of Agentic AI and its governance implications.

McKinsey

On 5 March 2026, The McKinsey Podcast featured a discussion between global editorial director Lucia Rahilly and partner Rich Isenberg on AI agents. According to Isenberg, the main problem for businesses is how to adopt AI agents safely, at scale, and with demonstrable return on investment. This points to the need for investment in governance, trust and AI literacy — unappealing yet foundational capabilities to ensure long-term value beyond initial gains.

KPMG

KPMG went a step further this month, by releasing “Agentic AI Gateway”, a best-practice standard dedicated to the use of AI agents in the automotive industry. The document highlights the role of identity and access management (IAM) as a critical control area, as agents may unintentionally leak sensitive information they have undue access to. KPMG therefore proposes a “Three C” framework, comprising Clarity (roles, responsibilities and oversight of AI agents), Control (providing the necessary access only), and Confidence (monitoring and auditing agents to enhance trust in the system). To operationalise this framework, organisations must consider the AI agent ecosystem (infrastructures, applications and stacks on which agents run), its control layer (effective IAM) and its data layer (data quality, usage and classification).

As we will see, the Agentic AI Gateway reflects a broader shift in the sector to consider IAM as foundational to agent governance.

Auriga

In April 2026, Indian IT consultancy Auriga published a white paper titled “Implementing Agentic AI: From Strategy to Scale”, including recommendations for deploying Agentic AI systems effectively. The report identifies five key risk management areas:

1. Production Gap: organisations often underestimate production demands, which is why many AI initiatives (including Agentic AI systems) do not progress beyond the pilot phase.

2. Guardrails: companies must build safeguards at the architectural level rather than at the prompt level. Prompt-level controls can be overridden; architectural controls cannot.

3. API Key Security: secret management, spending controls and IP whitelisting are essential to prevent unintentional misuse of APIs by agents.

4. Security & Data Governance: role-based access controls and least-privilege principles must be applied rigorously to prevent unintended use.

5. Regression Testing: monitoring and testing performance through validated regression suites are critical mitigating negative consequences of agent behavioural change.

What this signals

Consultancies clearly recognise the corporations’ growing need for Agentic AI governance frameworks. Similar to other industry actors, their approaches emphasise architectural design as the primary mechanism for ensuring safe and trustworthy Agentic AI deployment. The open question remains whether organisations are keeping pace with governance advances, or whether they disproportionately focus on short-term gains.

Financial services in focus

Many recent developments focus on financial services, a sector particularly sensitive to risks and opportunities of Agentic AI systems.

UK Finance and the Mills Review

In March 2026, UK Finance published its response to the *Mills Review*, an assessment of AI’s impact on the retail financial services industry, led by FCA’s executive director Sheldon Mills. The association highlighted the significant potential of AI agents, particularly in terms of optimising processes and automating decision-making.

At the same time, UK Finance raised concerns over:

· Liability. If actions are undertaken by multiple AI agents and errors cascade across the system, assigning responsibilities may become more complex.

· Human oversight. Human-in-the-loop mechanisms may reduce efficiency, suggesting a risk-based approach may be more appropriate when reviewing outputs.

· Transparency & Explainability. Evidencing how specific outcomes are generated may be challenging with multi-agent decision chains, creating challenges for accountability and consumer protection.

· Control effectiveness. More autonomous AI systems require firms to adapt their due diligence processes and demonstrate that control remain effective in dynamic systems.

Agentic AI in Private Equity Workflows

On 7 April 2026, AI consultant Leigh Coney (WorkWise Solutions) proposed the “Multi-Agent Orchestration Framework (MAOF)”, an Agentic AI system architecture designed to optimise private equity deal workflows. MAOF structures workflows by assigning responsibilities to specialised agents “with defined handoff protocols, structured verification checkpoints, and systematic human escalation”.

More white papers by Leigh Coney available here.

The framework is built around three core components:

1. Task decomposition. The deal workflow is divided into five sequential stages, each assigned to a dedicated agent “with a narrow context, a specific output schema, and defined success criteria”. These stages include: a) document ingestion and structure mapping; b) financial data extraction; c) risk flagging; d) thesis alignment scoring; e) memo synthesis. Each agent performs its task based on the output of preceding agent(s).

2. Confidence-based human escalation. Agent outputs are assigned a confidence score based on a three-band model: high (>0.85), medium (0.60 to 0.85), and low (<0.60). Confidence thresholds should be adjusted based on the firm’s error tolerance and the criticality of each task.

3. Self-correction loops with audit trails. Agents verify their own outputs against source documents before cascading them, while subsequent agents challenging the outputs received for source confirmation. The framework enables an immutable audit trail, with every decision logged (alongside its provenance). This supports both compliance continuous system improvement.

Agentic AI for retail banks

On 9 March 2026, the Boston Consulting Group (BCG) and OpenAI jointly released the report “How Retail Banks Can Put AI Agents to Work”, outlining practical applications of AI agents across banking operations. In particular, the report underscores significant efficiency gains offered by the automation of customer-facing and back-office processes. According to a BCG study, these enhancements could increase “banks’ profitability by 30% and reduce costs by 30% to 40% by 2030”.

The report describes three key areas for Agentic AI deployment:

· Customer credit onboarding. AI agents can “perform the initial analysis of the customer’s onboarding profile”, evaluating risk factors such as sanctions exposure, fraud indicators, and credit data. They can produce structured, traceable and reviewable risk assessments. Agents can augment the risk and compliance function rather than replacing it: for instance, they can expedite the review process, allowing human reviewers to focus on “discrepancies, exceptions, and higher-risk cases”.

· Back office. To deliver tangible impact, banks must evaluate agents against the same “document-heavy” and “exception-driven” tasks undertaken by human operators. This requires continuous measurement of “output quality, routing accuracy, exception rates, latency, and drift”. Critically, banks should establish a centralised control plane, that is an “in-house middleware layer” through which all AI systems operate. This layer would function as a front door, or an AI governance gateway, enabling consistent oversight and standardised controls.

· Governance mechanisms. The report advocates establishing an AI Centre of Excellence (CoE), namely a cross-functional team gathering technical, legal, compliance and risk experts to oversee AI systems holistically. The CoE is tasked with defining governance and risk frameworks, disseminating best practices, identifying use cases, and ensuring human oversight, auditability and alignment across control functions.

Agentic AI for financial reporting

On 16 April 2026, KPMG published a document outlining the risks posed by AI agents in financial reporting:

· Ungoverned agents: lack of comprehensive agent inventory, inclusive of accountability structures.

· Goal misalignment: agents may leverage internal loopholes, fail to consider the broader financial context, and thereby compromise reporting integrity.

· Cascading effects: agents may rationalise other agents’ inaccurate outputs, amplifying errors.

· Systemic risks: the impact posed by agents may propagate errors across interconnected IT systems.

· Segregation of Duties (SOD) conflicts: traditional SOD may be undermined by shared model components (e.g., a flaw in a shared ‘brain’ can potentially affect all agents).

To mitigate the above risks, KPMG proposes a set of workflow-level design interventions:

1. Strategic orchestration. Defining agents’ roles and permitted APIs and preventing unapproved agent-to-agent behaviour.

2. Human-centric oversight. Determining approval points, exception routing logic, and accountable owners.

3. Governance and accountability. Maintaining a central inventory to ensure clear ownership and risk classification.

4. Integrated security and compliance. Implementing least-privilege access controls, IAM, and continuous threat monitoring, aligned with regulatory and privacy requirements.

5. Rigorous testing and validation. Testing expected and unexpected adversarial scenarios, benchmarking outputs against baselines, and assessing “multi-agent interaction risk”.

6. Adaptive monitoring and continuous improvement. Monitoring performance, model drift, and anomalies, and iteratively retraining or fine-tuning based on observed behaviour.

Relatedly, on 14 April 2026, Kyndryl presented its Agentic AI solution for actuarial workflows, designed to address the supposed skill shortages in actuarial function while improving operational performance in insurance processes.

Converging design patterns

The rapid surfacing of Agentic AI frameworks in financial services suggest that financial operators are among the earliest and most significant short-term beneficiaries of Agentic systems. Automating high-risk tasks such as financial transactions and fraud detection requires thorough model evaluation, strong auditability assurances and strict compliance with applicable regulations. This necessitates architectural rather than procedural interventions.

Risk & control recommendations point towards a structural limitation: traditional governance mechanisms — those designed for static systems — struggle to scale as efficiently in dynamic agentic environments. Instead, governance seems to be shifting toward architectural re-design as primary mechanisms to address the complexity of agentic systems risks.

The Agentic Military AI Governance Framework

Governance of AI warfare systems is increasingly urgent yet remains underrepresented in the AI discourse. As such, it is commendable that, on 3 March 2026, Cambridge AI Safety Hub researcher Subramanyam Sahoo published a paper titled “The Controllability Trap: A Governance Framework for Military AI Agents”, proposing the Agentic Military AI Governance Framework (AMAGF).

The framework identifies six agentic failures:

1. Interpretive divergence (misalignment between operator intent and agents’ understanding)

2. Correction absorption (agents formally accept corrections while neutralising them)

3. Belief resistance (agents’ evidence-based judgment overrides operator authority)

4. Commitment irreversibility (cumulative minor tool calls cross irreversibility thresholds)

5. State divergence (operators’ mental model becomes incoherent with agent state)

6. Cascade severance (collective control loss through positive feedback loops)

The AMAGF architecture is divided into three governance pillars: a) prevention, aiming to lower the likelihood of control-failure; b) detection, aiming to identify degradation of control mechanisms over time; c) correction, aiming to restore controls. The framework includes a scenario-based illustration of its deployment.

World Economic Forum: The Agentic Readiness Framework

In April 2026, the World Economic Forum (WEF), in collaboration with Capgemini and innovation hub Berlin Global Government Technology Centre, released “Making Agentic AI Work for Government: A Readiness Framework”.

The framework is designed to support governments in assessing readiness for deploying Agentic AI systems and prioritising implementation at scale. Readiness assessments are structured around two key dimensions: system potential and implementation complexity. These are then categorised into three readiness tiers: high (early deployment), medium (phased implementation), and low (monitoring and long-term preparation).

Key recommendations from WEF comprise:

· Think in functions, not departments. Agents operate across organisational boundaries.

· Balance ambition with feasibility by accounting for implementation complexity before scaling.

· Start where best odds exist. Prioritise areas of highest readiness for initial deployment.

· Local context determines success. Global scores are baselines, but infrastructure, regulation, and culture determine feasibility.

· Expect the topography to evolve. The deployment landscape evolves continuously, requiring regular reassessment of readiness level.

Tech companies’ playbooks and toolkits

Tech companies have also contributed to shaping governance approaches.

IBM’s Agentic Playbook

This April, IBM published an article authored by its lead AI advocate Shalini Harkar, outlining the key components of a robust Agentic AI governance architecture. Harkar argues that the persistent gap between demo and production owes to the lack of appropriate governance structures require to scale agentic systems effectively. Scaling agentic systems, she writes, “requires aligning value, governance, architecture and people in a single operating model”.

First, companies need to establish five core layers of control: (i) define the agents’ scope, (ii) set system boundaries, (iii) pre-deployment validation, (iv) monitoring, and (v) regular review.

Key governance trade-offs prior to deployment:

1. Calibrating the right trade-off between speed of and control in agent deployment

2. Balancing innovation with predictability, especially from a governance perspective

3. Establishing accountability structures and escalation paths in the event of agent failure

4. Designing controls that scale relative to the increased agent autonomy and reach

5. Introducing mechanisms preventing increased automation from eroding user trust

Finally, the framework advocates a governance-by-design approach in the post-deployment phase, grounded in operational clarity across ownership, authority, decision-making, and control boundaries.

Microsoft’s Agent Governance Toolkit

In the same month, Microsoft Principal Group Engineering Manager Imran Siddique published the “Agent Governance Toolkit”, an open-source project released “under the Microsoft organization and MIT license that brings runtime security governance to autonomous AI agents”. It consists of a seven-package toolkit, available in Python, TypeScript, Rust, Go, and .NET.

AWS’ Multi-Agent Agentic AI Security Framework

On 7 March 2026, AWS released “Securing Multi-Agent Agentic AI Systems With Design Principles and Prioritization Framework”. The document provides a four-tier framework (ranging from no agency to prescribed, supervised and full agency), each mapped to corresponding risk levels. AWS recommends a progressive implementation, starting with low-risk tiers to ensure that higher-autonomy systems are evaluated against their capabilities, risk and auditing requirements.

Building on this model, the report proposes five design principles:

1. Progressive Autonomy. Implementing progressive autonomy with bounded agency: begin with minimal agency autonomy and expand as security controls mature, enforcing explicit boundaries at each stage.

2. Continuous Monitoring. Establish continuous monitoring and behavioural validation through anomaly detection, decision-pattern analysis, and scope creep.

3. Human Oversight. Preserve human oversight through intervention points, including override capabilities and approval workflows designed to avoid reviewer overload.

4. Identity & Authorisation. Secure identity, authorisation and tool access using least-privilege IAM controls across APIs, tools, and integrations to constrain agent scope.

5. Explainability. Design for explainability and auditability by embedding transparency-by-design mechanisms and automated logging of agent decision and rationale.

AWS further decomposes agentic security into six critical operational dimensions:

· Identify context. AWS prioritises authentication and authorisation, highlighting the “confused deputy problem”, where lower-privilege entities “elevate permissions through agents”.

· Threat modelling. Also a critical dimension, threat modelling and risk assessments align with zero-trust security principles.

· Audit & logging. These are treated as immediate requirements at deployment.

· Memory & state protection. Data memory and state protection through session isolation and monitoring for hallucination propagation and cross-agent contaminations.

· Model controls. Agentic and foundation model controls, inclusive of validation layers and filtering of PII, sensitive data and profanities where applicable.

· Agency perimeters. Agency perimeters and policies through tool registries, security assessments, and RBAC enforcement for tool usage.

Futurum’s Agentic Enterprise Governance Model

In April 2026, US-based tech advisory Futurum released a report titled “Gemini Enterprise: Governing and Scaling the Agentic Enterprise”, examining the opportunities offered by Gemini Enterprise in building and managing autonomous agents and generative AI workflows.

According to Futurum, the main barrier to enterprise Agentic AI scaling lies in the lack of the necessary management infrastructure to govern agent population at production scale:

· Governance and Visibility. Lacking complete visibility over what agents are doing raises substantially operational risks. Key concerns noted by leaders include security and data privacy, loss of human control, and regulatory and compliance risks.

· Integration and Context. Organisations need to bridge the gap between agents limited to a single session and agents operating across application workflows. This involves enhancing memory retention and cross-boundary operation as pre-requisites for implementing Agentic AI at scale.

· Observability and Evaluation. Another obstacle to production at scale is the lack of a robust, deterministic model evaluation to ensure continued high performance. In this context, the so-called semantic layer has established itself as mission-critical layer for strengthening trust in AI agents.

These constraints translate into five foundational capabilities required for scalable governance:

1. Unified governance and identity layer. Each agent should have a defined identity recorded in centralised registry.

2. Persistent memory and long-context support. Workflows rarely complete in a single session: agents need to be given access to enterprise-wide data while retaining context across interactions.

3. Broad model and integration flexibility. A mature and effective agentic platform should integrate with existing systems without introducing unnecessary dependencies.

4. A coherent path from no-code to full-code development. Agent deployment should provide a seamless path from visual to full-code development to reduce friction from design to production.

5. Observability, simulation and evaluation. Operational agentic governance requires evaluating agent behaviour pre-production, tracing reasoning during operation, and iteratively improving performance.

Futurum effectively frames Agentic AI governance as a platform problem rather than a policy problem.

Tech firms’ responsibility

It is notable that some tech firms are providing sophisticated governance and control frameworks and seemingly keeping pace with the increasing capabilities of AI agents. These contributions demonstrate that integrating technically robust control measures with strategic business objectives is not only feasible, but necessary for deployment at scale.

As autonomous systems become more deeply embedded within business processes, the boundary between technical and governance design grows progressively thinner. In these environments, the gap between system capabilities, human oversight and accountability is likely to widen. That is, unless it is explicitly bridged through architecture-level controls.

The availability of governance mechanisms such as those outlined above removes the justification for companies to implement risk-blind deployment. Deploying without appropriate safeguards and control measures therefore becomes a question of organisational choice rather than a technical limitation. In this context, social responsibility becomes a core design constraint for Agentic AI systems.

Academic Papers

Many academic and industry contributions treat governance as a runtime function instead of a mere policy layer. Below is a brief overview of some of these works.

AICEBERG

On 22 April 2026, cybersecurity experts Florin Popescu and Denis Stanescu published an article in an MDPI journal proposing an Agentic AI framework for “radio monitoring, compliance and governance based on LLM, MCP, and SCPI in Smart Cities”, called AICEBERG. Developed for the purpose of enhancing urban radio spectrum monitoring, the proposed framework aims to reduce “human dependency, enhance reproducibility, and lower the expertise barrier required for RF spectrum surveillance”. Learn more about AICEBERG.

Ethical challenges in Agentic AI systems

Last February, Birupaksha Biswas and Suhena Sarkar (from the Burdwan and Kolkata Medical Colleges, respectively) authored a paper addressing key limitations in current approaches to Agentic AI governance. These comprise “checking goal congruency, tracking on emergent capabilities, engineering against negative instrumental actions, and accountability in multi-agent systems”. The paper is a comprehensive review of risks and ethical challenges associated with AI agents, suitable for both technical and non-technical audiences. Read full paper.

The FASTRAC Framework

On 22 April 2026, Syed Muhammad Khuzaima Alam, a student at the University of Illinois, proposed the “FASTRAC” framework for AI agents executing financial transactions in B2B environments. FASTRAC (Financial Agent Safety, Trust, Risk, and Compliance) introduces a governance function integrating “constraints, validation, risk scoring, monitoring, and learning into a closed-loop system”. By defining execution as a “constrained optimization problem under risk thresholds”, FASTRAC ensures that agent decisions remain compliant, auditable, and controllable. Read more about FASTRAC.

Arbiter-K

Arbiter-K is the name of a governance-first architecture proposed by 12 Chinese and Hong Kong researchers in a paper published on 20 April 2026. The paper argues that the main challenge in moving from pilot to production lies in the prevailing orchestration paradigm, where the LLM runs the loop, and developers add guardrails afterwards. The researchers invert this approach by embedding the LLM in a controlled execution system rather holding it as the brain and controller. In this design, the LLM suggests, the system decides and enforces. This is by far one of the most technically detailed frameworks in this review, and is particularly relevant to readers familiar with computer architecture language. Discover Arbiter-K.

Agentic AI Governance for Fraud Detection

New Jersey-based AI researcher Jalendar Reddy Maligireddy authored an Agentic AI framework designed to govern real-time fraud detection. The framework is hinged on three layers: perception (behavioural biometrics and transaction telemetry ingestion), reasoning (combining graph neural networks with LLM-based anomaly reasoning), and governance (HITL oversight, XAI, automated compliance reporting). The framework reportedly demonstrated promising results in scenario-based evaluations. Find out about this framework.

Closing the governance gap

On 18 April 2026, independent researchers Christopher Koch and Joshua Wellbrock proposed three solutions to address gaps in Agentic AI governance literature, namely: (i) a four layer framework (evaluation, governance, orchestration, and assurance); (ii) an ODTA (Observability, Decidability, Timeliness, Attestability) runtime-placement test; (iii) a minimum action-evidence bundle for state-changing actions. Read full paper.

An Agentic AI Security Architecture

Last February, five researchers from the US-based Northeastern University introduced SAGA, a security architecture for governing Agentic AI systems. The framework introduces the “Provider”, a central entity maintaining agent information and access control policies, alongside a cryptographic mechanism for generating access control tokens, to enable fine-grained control over agent interactions. Learn more about SAGA.

Explainable Agentic AI

PhD and Credo AI senior researchers Yomna Elsayed and Cecily K Jones authored a paper addressing the enterprise need for explainability mechanisms to support agent deployment at scale. The researchers identified the following measures as essential for trustworthy Agentic AI: agent inventory, agent cards and dependency graphs (design phase), and deep observability, contextual traceability and operational monitoring (runtime phase). Learn more about explainable Agentic AI.

The Auton Agentic AI Framework

In February, a group of researchers at Snapchat outlined the Auton Agentic AI Framework, an architecture to standardise the creation, execution, and governance of autonomous agent systems. The framework separates the “cognitive blueprint” (a language-agnostic specification of an agent’s identity and capabilities) and “runtime engine” (a platform-specific execution substrate that instantiates and runs the agent). This separation, according to the researchers, enables “cross-language portability, formal auditability, and modular tool integration via the Model Context Protocol”. Read full white paper.

The bigger picture

The proliferation of frameworks and use cases of AI agents in the above academic contributions is closely aligned with the industry and consultancy frameworks discussed earlier. In fact, much like other models, these papers also address the need for architecture-level governance. These approaches also converge on the idea that agentic systems should be treated as distributed and interactive systems requiring constraints and oversight.

Collectively, these contributions suggest that scalability of Agentic AI depends less on the agentic capabilities per se and more on the maturity of the governance infrastructure that surrounds and supports it. Researchers and industry experts are acutely aware of this shift and demonstrate that these risks are being actively addressed, including in niche use cases, where societal stakes are comparatively lower than in other domains.

Overall, the Agentic AI ecosystem appears more receptive to the need for governance and controls than the traditional AI. Academia, consultancies, and industry experts have displayed substantive efforts in designing governance and control architectures for AI agents.

As companies struggle to move from Agentic AI pilot to production, the reviewed approaches outline how organisations can deploy agents in a way that ensures long-term sustainability: safely for individuals and in compliance with applicable regulations. As always, however, whether these frameworks will be implemented is a decision that ultimately rests with business leaders. This is likely to become a key differentiator between organisations that can scale Agentic AI systems responsibly and those that cannot.

Sources cited

Consultancies

· The McKinsey Podcast with Lucia Rahilly and Rich Isenberg

· Implementing Agentic AI: From Strategy to Scale, Auriga IT

· Agentic AI Gateway, KPMG

· Gemini Enterprise: Governing and Scaling the Agentic Enterprise, Futurum

Financial services

· UK Finance response to: Review into the long-term impact of AI on retail financial services (The Mills Review)

· Agentic AI in Private Equity: Multi-Agent Orchestration for End-to-End Deal Workflows, Leigh Coney

· How Retail Banks Can Put AI Agents to Work, BCG & OpenAI

· *From Branches to Bots: Will AI Agents Transform Retail Banking?*, BCG

· Agentic AI workflows in financial reporting, KPMG

· Agentic AI solution for actuarial workflows, Kyndryl

Military

· The Controllability Trap: A Governance Framework for Military AI Agents, Subramanyam Sahoo

International Organisations

· Making Agentic AI Work for Government: A Readiness Framework, WEF

Tech Companies

· Agentic AI governance–Playbook, IBM

· Agent Governance Toolkit: Open-source runtime security for AI agents, Microsoft

· Securing Multi-Agent Agentic AI Systems With Design Principles and Prioritization Framework, Amazon

Academic Papers

· AICEBERG: A Novel Agentic AI Framework for Autonomous Radio Monitoring, Compliance and Governance Based on LLM, MCP, and SCPI in Smart Cities (Florin Popescu, Denis Stanescu)

· Responsible agentic artificial intelligence governance: Risk, safety, and ethical challenges in autonomous systems (Birupaksha Biswas, Suhena Sarkar)

· Governing the Agentic Enterprise Guardrails for Autonomous AI Agents in B2B (Syed Muhammad Khuzaima Alam)

· From Craft to Kernel: A Governance-First Execution Architecture and Semantic ISA for Agentic Computers (Xiangyu Wen, Yuang Zhao, Xiaoyu Xu, Lingjun Chen, Changran Xu, Shu Chi, Jianrong Ding, Zeju Li, Haomin Li, Li Jiang, Fangxin Liu, Qiang Xu)

· Agentic AI Governance Framework for Real-Time Fraud Detection in Digital Payment Systems: A Multi-Layered Architecture for Financial Security, Jalendar Reddy Maligireddy

· Beyond Task Success: An Evidence-Synthesis Framework for Evaluating, Governing, and Orchestrating Agentic AI (Christopher Koch, Joshua A. Wellbrock)

· SAGA: A Security Architecture for Governing AI Agentic Systems (Georgios Syros, Anshuman Suri, Jacob Ginesin, Cristina Nita-Rotaru, Alina Oprea)

· Agentic Explainability at Scale: Between Corporate Fears and XAI Needs XAI at Scale (Yomna Elsayed, Cecily K Jones)

· The Auton Agentic AI Framework. A Declarative Architecture for Specification, Governance, and Runtime Execution of Autonomous Agent Systems (Sheng Cao, Zhao Chang, Chang Li, Hannan Li, Liyao Fu, Ji Tang)