1. Introduction: The end of the "one-brain" illusion

In recent years, enterprise AI has been dominated by a single promise: that one comprehensive platform could automate every cognitive process across a financial institution including risk, compliance, customer service, and operations.

The idea is appealing. For banks constrained by fragmented data and rising costs, the vision of a unified intelligence core suggests simplicity and efficiency. But the institutional logic of finance runs in the opposite direction. Financial systems are built on segregation of authority, traceability of reasoning, and regulatory accountability. Concentrating all decision-making inside a single opaque model conflicts with those foundations.

Evidence from the field reinforces this constraint. Even as major vendors launch integrated AI suites, early adopters in banking are moving toward distributed agent architectures: specialized, explainable systems coordinated through governance layers rather than absorbed into one brain. BNY Mellon's use of domain-specific digital workers (WSJ 2025) and multi-agent orchestration pilots observed by FinRegLab (2025) illustrate this shift.

The future of institutional AI is unlikely to be monolithic. It is emerging as a federated ecosystem of agents, interoperable through open protocols, governed by orchestration fabrics, and accountable under evolving regulatory norms.

2. The signals of change

Several developments in 2025 point toward this multi-agent model:

Open interoperability protocols

The Linux Foundation's Agent2Agent (A2A) Protocol, launched in June 2025, enables secure communication among AI agents across vendors (Linux Foundation 2025). Backed by Microsoft, AWS, Anthropic, and Hugging Face, it represents the first coordinated effort to create a shared language for autonomous systems. In parallel, the Model Context Protocol (MCP), introduced by Anthropic and now adopted by AWS and others, defines standardized interfaces for tool access and memory management (AWS Open Source Blog 2025). Together, these protocols begin to do for intelligence what HTTP and TCP/IP did for information exchange.

Enterprise adoption signals

The Wall Street Journal reported that BNY Mellon and other major institutions are already deploying digital workers: function-specific agents performing reconciliation and compliance tasks (WSJ 2025). These are limited in scope but demonstrate how agentic systems can coexist with legacy infrastructure.

Financial-sector research

A September 2025 study by FinRegLab described agentic AI as "the next wave arriving in financial decisioning" but emphasized persistent challenges around governance, explainability, and integration (FinRegLab 2025).

Capability constraints

According to EY's 2025 Banking AI Report, 58 percent of financial institutions cite technology-skills shortages as their main obstacle to scaling AI (EY 2025). This explains why most organizations will rely on hybrid sourcing models that combine internally developed agents with vendor solutions connected through open standards.

These are signals of experimentation rather than maturity, but together they mark a structural shift.

3. From tools to digital labor

AI agents differ from earlier automation tools because they perform tasks that resemble cognitive labor: assessing credit, detecting fraud, generating reports, or advising clients. They operate semi-autonomously under human oversight, governed by parameters such as cost, accuracy, and compliance.

In economic terms, agents represent units of productive capability. Institutions will increasingly evaluate them not by ownership or license cost but by decision quality per unit cost: how efficiently and reliably each agent converts data into compliant, actionable outcomes.

This reframes intelligence as an economic resource rather than a fixed asset, subject to competition, pricing, and regulation.

4. The role of open protocols

The importance of A2A and MCP lies in their ability to enable interoperability without consolidation. They create shared formats for discovery, authentication, and data exchange among agents from different vendors.

In practice:

  1. Each agent exposes an Agent Card describing its capabilities and trust credentials (Aisera 2025).
  2. Institutional orchestration layers route tasks between agents according to policy and permission.
  3. The protocol manages negotiation, error handling, and logging.
  4. All interactions are cryptographically verifiable, ensuring audit trails.

This allows institutions to combine internal and external agents safely. For example, a compliance agent developed in-house can query a risk-assessment agent provided by a vendor through the same protocol interface.

However, interoperability remains incomplete. Competing standards, including the W3C's Agent Protocol initiative, may fragment adoption, and large-scale production testing is still limited.

5. The orchestration fabric: governance as infrastructure

Inside a regulated institution, multiple agents cannot function without a governance and control layer known as the Agent Orchestration Fabric. This layer provides:

  • Identity management for every agent and model
  • Policy enforcement defining data and action permissions
  • Routing and coordination among agents and human supervisors
  • Logging and explainability for audit and retraining
  • Lifecycle management for deployment, suspension, and retirement

Conceptually, this fabric serves as the institutional nervous system for digital labor. It transforms autonomy into governed behavior and is essential for regulatory compliance.

6. Build versus buy: an emerging decision matrix

The strategic question for most institutions is not whether to use agents, but when to build them and when to buy them.

Internal development (Build)

Institutions tend to build agents when they depend on proprietary data or decision logic that defines competitive advantage, when operating in high-stakes domains such as risk and compliance where explainability is mandatory, when integration with legacy systems requires institutional knowledge, or when they want to minimize vendor lock-in and retain control.

External procurement (Buy)

Institutions typically buy agents when the function is non-differentiating, when speed to value outweighs customization, when internal AI skills are insufficient (EY 2025), or when vendors provide certified interoperability through A2A or MCP.

Hybrid reality

Most financial organizations will adopt both strategies. They will build sovereign agents for core decision domains and integrate external ones for peripheral or fast-moving tasks. Open protocols make this mix technically feasible and governance compatible.

7. Cooperation and competition among agents

Open standards create a new dynamic of cooperative competition. Agents from different providers can interact while still competing on performance, cost, and compliance.

  • Cooperation allows agents to share context and delegate subtasks across vendor boundaries, improving overall efficiency.
  • Competition ensures each agent is selected based on reliability and efficiency, introducing a market-like mechanism for intelligence.
  • Over time, performance benchmarking could enable marketplaces that publish standardized metrics for agent accuracy and trustworthiness, although such markets remain hypothetical today.

The result is a gradual emergence of intelligence liquidity, where decision-making capacity can be reallocated much like financial capital, though practical and regulatory constraints remain substantial.

8. Constraints and uncertainties

Despite the momentum, major uncertainties persist:

Domain Key Challenge Evidence / Observation
Interoperability Competing standards (A2A, MCP, W3C) risk fragmentation Enterprise pilots remain vendor-specific (AWS 2025)
Governance and liability Lack of defined regulatory frameworks for autonomous agents FinRegLab (2025) notes absence of accountability mechanisms
Security surface Agent credentials and API permissions expand attack vectors CyberArk (2025) identifies "machine identity sprawl"
Data boundaries Persistent memory and cross-agent context raise compliance issues Microsoft (2025) highlights debate over memory retention
Legacy integration Core banking infrastructure limits real-time orchestration Everest Group (2025) cites integration as a major barrier

The trajectory toward fully interoperable agent ecosystems will likely be gradual and path dependent.

9. Strategic implications

For financial institutions

  • Build governance infrastructure before scaling agent use.
  • Engage with open-standard alliances to influence interoperability norms.
  • Map capabilities to determine where intelligence should be built or sourced.
  • Certify and monitor agents as auditable digital employees.
  • Redefine efficiency metrics around decision outcomes, not headcount reduction.

For vendors

  • Adopt open protocols early to remain compatible with institutional architectures.
  • Focus on specialization and explainability as differentiators.
  • Provide transparent performance metrics to enable benchmarking.

For regulators

  • Recognize agents as semi-autonomous actors requiring identity and accountability frameworks.
  • Encourage interoperability and transparency to prevent market capture.
  • Move from post-event oversight to continuous verification through standardized logging.

10. Probable trajectory

Phase Timeline Characteristics
Pilot experimentation 2025–26 Limited multi-agent pilots in compliance, customer service, and risk
Governance fabric formation 2027–28 Institutions formalize orchestration and monitoring layers
Protocol standardization 2028–29 A2A and MCP gain maturity; interoperability certifications emerge
Market integration 2030+ Agents benchmarked as service units; governance and regulation mature

11. Conclusion: Evidence over ideology

The evidence from 2025 suggests that the agentic model of AI is moving from concept to cautious implementation in finance. Open protocols such as A2A and MCP are providing the technical foundation for interoperability, while institutions like BNY Mellon and FinRegLab demonstrate early operational use.

However, fragmentation, security risk, and uncertain regulation remain significant constraints. Whether multi-agent systems become dominant will depend less on technological capability and more on how governance, trust, and interoperability mature.

Intelligence in finance is beginning to behave like an economic market: distributed, competitive, and governed by trust. The rational path forward for institutions is not hype or hesitation, but structured experimentation grounded in evidence and open standards.

References

  1. Aisera. "Agent Cards: Standardizing AI Agent Capabilities." Technical Documentation, 2025.
  2. AWS Open Source Blog. "Model Context Protocol Adoption in Enterprise AI." AWS Blog, 2025.
  3. CyberArk. "Machine Identity Sprawl: The New Security Challenge." Security Research Report, 2025.
  4. EY. "2025 Banking AI Report: Technology Skills and Scaling Challenges." EY Global Financial Services, 2025.
  5. Everest Group. "Legacy Integration Barriers in AI Agent Deployment." Industry Analysis, 2025.
  6. FinRegLab. "Agentic AI in Financial Decisioning: Governance and Integration Challenges." Research Study, September 2025.
  7. Linux Foundation. "Agent2Agent (A2A) Protocol: Enabling Interoperable AI Agents." Technical Specification, June 2025.
  8. Microsoft. "Memory Retention and Cross-Agent Context in Financial AI." Technical Whitepaper, 2025.
  9. Wall Street Journal. "BNY Mellon Deploys Digital Workers for Financial Operations." WSJ Technology Section, 2025.