The conversation around artificial intelligence in the enterprise has fundamentally shifted. For the past two years, the dominant narrative centred on copilots and generative assistants, new digital tools that suggest code, draft emails, or summarise documents while a human user remains firmly in the driver’s seat. These tools have been quite useful, but their drawbacks are now really obvious in more complicated business settings. They need frequent prompts, have no memory of past interactions, can’t come up with their own plans, and simply stop functioning the moment humans pause or step away.
In 2026, the era of the copilot is giving way to the era of the autonomous agent. Open agent systems are moving beyond simple text generation to active, goal-driven execution. Instead of answering questions, these systems get work done. They perceive their environment, break down complex goals into actionable sub-tasks, interact with external tools and APIs, collaborate with other agents, and evaluate their own progress. This article explores the architecture of modern open agent systems, the open standards making enterprise-scale deployment possible, their practical use cases across industries, and the governance challenges that stand between experimentation and production.
The Architecture of Autonomy: Moving Beyond the Chatbot
To understand why agentic AI is genuinely transformative, one must first understand how it differs from traditional large language models (LLMs). An LLM is essentially a highly sophisticated statistical engine that maps patterns in text. An AI agent, conversely, wraps that intelligence core inside a systematic framework capable of executing complex, multi-step loops. The distinction is not cosmetic, it reflects a fundamental change in how work gets initiated, executed, and verified.
Modern open agent systems rely on four core architectural components to achieve autonomy. The first is perception and context awareness, which allows the agent to ingest and interpret real-time data from multiple sources, such as electronic health records, enterprise databases, or sensor streams, rather than responding to a single static prompt.
The second is reasoning and planning, where the agent decomposes high-level directives into manageable steps, adjusting its strategy dynamically if obstacles arise. The third is tool use, the critical capability that allows agents to read and write to repositories, execute code, query databases, and interact with external software applications. The fourth is persistent memory and continuous learning, which enables agents to adapt their behaviour based on past outcomes, evolving from static tools into adaptive digital teammates.
The transition from single-agent frameworks to multi-agent ecosystems represents the most significant architectural leap of 2026. Instead of relying on a monolithic model to handle all aspects of a task, modern systems deploy specialised agents that collaborate. A software development workflow, for instance, might involve a researcher agent, a coding agent, a testing agent, and a reviewer agent, all sharing context and coordinating their efforts to complete a project.
This modular approach enhances reliability, scalability, and security, allowing enterprises to isolate functions and apply specific guardrails to individual agents rather than managing a single, opaque system.
The market has formalised this architecture rapidly. The average company already runs twelve AI agents today, a number expected to reach twenty by 2027, according to Salesforce’s 2026 Connectivity Report [1]. However, half of those agents currently operate in complete isolation (no agent-to-agent communication, no shared context, no handoffs) which significantly limits what multi-agent systems can accomplish and underscores the urgency of the standardisation work now underway [1].
5 Types of AI Agents: Autonomous Functions & Real-World Applications
The Standardisation Imperative: Protocols Driving the Agentic Web
The rapid proliferation of AI agents in late 2024 and 2025 created a severe integration bottleneck. Enterprises found themselves managing hundreds of custom integrations to connect their agents to various tools and to each other. Before open standards, connecting five agents to ten tools meant up to fifty bespoke integrations, each one a maintenance liability [6].
In 2026, this fragmentation is being resolved through the widespread adoption of open protocol standards, creating a robust infrastructure for what practitioners are calling the “agentic web.”
Two foundational protocols have emerged as the bedrock of enterprise agent deployments: the Model Context Protocol (MCP) and the Agent-to-Agent (A2A) protocol. They are complementary, not competing, and a serious enterprise deployment runs both.
| Protocol | Primary Function | Governance | Status in June 2026 |
| Model Context Protocol (MCP) | Standardises how a single agent connects to external tools, data sources, and APIs (the “vertical” problem). | Open-source, Anthropic-stewarded | De facto standard; thousands of active servers; adopted by OpenAI, Microsoft, and Google DeepMind [6]. |
| Agent-to-Agent (A2A) | Standardises how independent agents discover, communicate, and collaborate with one another across different platforms and vendors (the “horizontal” problem). | Linux Foundation (donated mid-2025) | Production-grade; 150+ supporting organisations; integrated into Microsoft Azure AI Foundry, AWS Bedrock AgentCore, and Google Cloud [6]. |
MCP equips an agent with capabilities; A2A provides the shared language for agents to collaborate with peers they did not build. Anthropic released MCP in November 2024 as a standard for connecting AI assistants to the systems where data lives. Google introduced A2A in April 2025 and donated it to the Linux Foundation shortly thereafter.
By its April 2026 one-year milestone, A2A had crossed 150 supporting organisations and reached production-grade enterprise deployments across supply chain, financial services, insurance, and IT operations, with founding partners including Amazon Web Services, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow.
However, as agents gain autonomy, identity and trust become paramount concerns that neither MCP nor A2A fully addresses. To fill this gap, the Linux Foundation announced the intent to launch the Agent Name Service (ANS) on 23 June 2026 [7]. Built on the existing Domain Name System (DNS) infrastructure, ANS provides a federated framework for securely identifying and verifying autonomous agents at internet scale [8].
This allows organisations to publish agent identities through domains they already control, enabling other systems to verify an agent’s ownership, permissions, and operational history before interacting with it [7].
These three open standards – MCP, A2A, and ANS – are transforming agentic AI from isolated experiments into interoperable, governable enterprise infrastructure. The protocols are now settled. The interesting decisions have moved up a layer, to orchestration, governance, and measurable business outcomes.
Building Trust in the AI Era: Agent-to-Agent Communication With DIDs and VCs – Alexander Shcherbakov
Practical Enterprise Use Cases in 2026
The theoretical promise of agentic AI is rapidly translating into measurable business value across multiple sectors. According to Gartner, 40% of enterprise applications will feature task-specific AI agents by the end of 2026, a massive increase from less than 5% in 2025, an eightfold jump in a single year that outpaces the early adoption curves of both cloud and mobile computing.
Transforming Healthcare Delivery and Operations
The healthcare industry, burdened by administrative overhead and clinician burnout, is aggressively adopting agentic AI. A February 2026 Deloitte survey of 100 US healthcare technology executives revealed that 61% of healthcare organisations are already building or budgeting for agentic AI, with 85% planning to scale up investments over the next two to three years [11].
How AI Is Reshaping Global Health: From Precision Surgery To Food Security | Raising Expectations
Agentic AI in healthcare operates across three primary domains. In clinical decision support, agents move beyond static recommendations to simulate treatment paths tailored to a patient’s unique physiology and genomics, updating dynamically as new lab results or vital signs emerge [3].
In care coordination, agents autonomously manage patient transitions (scheduling home health visits after discharge, coordinating follow-up calls, and updating electronic health records (EHRs) in real time) closing the loop between sensing and doing without waiting for human input.
Generative AI Meets the Healthcare System: Key Lessons from the Early Years | 11 Feb 2026
In administrative automation, agents are streamlining utilisation management, prior authorisation processes, and claims management, with more than 70% of health plans prioritising these workflows for agentic deployment.
A concrete example of this shift is Humana Inc.’s member support operation. The company launched an agentic AI tool to help its member advocates respond to member questions with greater speed and precision. The tool summarises call conversations, anticipates member needs, and autonomously guides advocates with the most relevant information, allowing human staff to focus on care and engagement rather than information retrieval.
Revolutionising Software Engineering
Software development is experiencing a profound shift from AI-assisted coding to agentic engineering. According to Anthropic’s 2026 Agentic Coding Trends Report, developers now use AI for roughly 60% of their work, yet can only fully delegate 0–20% of tasks to agents without oversight [1].
McKinsey reports that while 62% of organisations are experimenting with AI agents in software development, fewer than 25% have scaled them to production, a 37-point gap that reflects not a capability problem but an operational and organisational one.
The most successful deployments are occurring in high-verifiability domains, where the output of an agent can be checked quickly and objectively. Code review agents autonomously analyse pull requests for regressions, style violations, and security vulnerabilities, while test generation agents produce comprehensive test suites from code diffs.
Operations teams are deploying agents for incident response, where they can autonomously triage alerts, surface relevant runbooks, and draft initial remediation steps, drastically reducing the cognitive load during critical outages.
As these agents mature, they are moving beyond execution tasks to manage institutional knowledge, autonomously generating documentation and onboarding materials based on codebase changes, addressing one of the most consistently skipped tasks in engineering without asking developers to change their workflow.
Anthropic’s Boris Cherny: Why Coding Is Solved, and What Comes Next
Orchestrating the Global Supply Chain
In logistics and supply chain management, agentic AI is bridging the critical gap between visibility and action. Traditional supply chain control towers provide dashboards that monitor and answer “what is happening,” but they require human analysts to interpret the data and coordinate responses across disconnected systems.
The result is delayed execution at enterprise scale: inventory issues surface after impact occurs, supplier risks remain buried in disconnected workflows, and teams spend more time interpreting reports than acting on them.
Agentic supply chain systems operate on a fundamentally different logic. They continuously monitor enterprise data, understand the relationships between various signals, and execute actions autonomously. For example, if a supplier delay is detected, an agent can autonomously assess the impact on inventory, evaluate alternative sourcing options, and initiate rebooking procedures, all while providing human operators with transparent reasoning and traceable SQL logic for its decisions.
Oracle’s AI Supply Chain Tower, built using OCI native services and MCP-based connectivity, demonstrated this capability by assembling a fully functioning operational system in under eight hours using existing enterprise assets, without a new data warehouse or lengthy implementation cycle [12]. This shift from passive reporting to active orchestration is critical for managing the volatility of modern global supply chains.
Powering Supply Chain With Agentic AI (By McKinsey)
The Production Gap: Navigating the Challenges of Scale
Despite the rapid adoption and clear use cases, scaling agentic AI from pilot to production remains the defining challenge of 2026. A striking gap exists between experimentation and production: while 88% of organisations report using AI in at least one function, only 7% have fully scaled it across the enterprise. Gartner’s analysis is sobering: over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls [13].
Most agentic AI propositions currently lack significant return on investment, as models do not yet have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time.
The primary barrier to scale is not a lack of capability, but a lack of robust governance and operational infrastructure. When organisations move from pilots to production, they encounter three critical gaps that neither MCP nor A2A resolves.
The first is governance and policy: protocols define how agents exchange messages, but they do not define whether a finance agent is permitted to approve a euro payment without human review, or which agents can invoke which others. The second is observability: when a five-agent workflow produces a wrong answer, no protocol identifies which agent reasoned badly, which tool returned stale data, or where the chain broke [6]. The third is cross-agent identity and authorisation: signed Agent Cards verify that an agent is who it claims to be, but they do not manage what that agent is permitted to do on behalf of which human, with what credentials, and under what audit trail.
To bridge this production gap, enterprises must adopt what Persistent Systems describes as an “AI operating system” approach: centralising repeatable platform work such as identity management, policy enforcement, logging, evaluation, and cost controls, rather than rebuilding these foundations for every new agent deployment [14].
The organisations that succeed in 2026 are those that treat agentic AI not as an innovation lab experiment, but as a core software engineering discipline, with governance, observability, and measurable business outcomes built in from day one.
The transition from generative AI to agentic AI represents a fundamental reorientation of enterprise technology. Open agent systems, empowered by robust architectural frameworks and standardised protocols like MCP, A2A, and ANS, are moving beyond digital assistance to become autonomous drivers of business value.
From healthcare and software engineering to finance and global supply chains, these systems are executing complex workflows, adapting to dynamic environments, and redefining the nature of work itself.
However, the path to agentic automation is not without friction. The high projected failure rate of early projects underscores the critical need for rigorous governance, transparent observability, and a strategic approach to deployment. The protocols are settled. The infrastructure is maturing. What separates the organisations that will lead from those that will merely follow is not the decision to adopt open agent systems, (that decision is increasingly being made for them by market forces) but the discipline with which they govern, observe, and scale them. In 2026, the competitive advantage belongs not to those who move fastest, but to those who move most deliberately.
What are your thoughts on the steps organizations should take now to establish a strong foundation for scalable, well-governed, and effective agentic AI deployment in the future?
References
Agrawal, K. (2026). Agentic AI / AI Agents: Future of Enterprise Automation in 2026. [online] Available at: https://kanishkagrawal.com/agentic-ai-ai-agents-enterprise-automation-2026/ [Accessed 11 July 2026].
Alice Labs (2026 ). AI Agent Frameworks 2026: Production-Tested Ranking. [online] Available at: https://alicelabs.ai/en/insights/best-ai-agent-frameworks-2026 [Accessed 9 July 2026].
Appinventiv (2026 ). How Agentic AI in Healthcare Is Bringing in Industry-level Transformation. [online] 22 June. Available at: https://appinventiv.com/blog/agentic-ai-in-healthcare/ [Accessed 8 July 2026].
Beam AI (2026 ). Agent2Agent vs MCP: The 2 Protocols Your 2026 AI Agent Stack Actually Runs On. [online] 23 June. Available at: https://beam.ai/agentic-insights/agent2agent-vs-mcp-2026-ai-agent-stack [Accessed 14 July 2026].
Deloitte (2026 ). Many health care leaders are leaning into agentic AI as adoption hurdles ease. [online] 11 February. Available at: https://www.deloitte.com/us/en/insights/industry/health-care/agentic-ai-health-care-operating-model-change.html [Accessed 15 July 2026].
Gartner (2025a ). Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025. [online] 26 August. Available at: https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025 [Accessed 15 July 2026].
Gartner (2025b ). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. [online] 25 June. Available at: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 [Accessed 14 July 2026].
Growin (2026 ). AI Agents in Software Development: A 2026 CTO Guide. [online] 29 June. Available at: https://www.growin.com/blog/ai-agents-in-software-development-26/ [Accessed 15 July 2026].
Infoworld (2026 ). New Linux Foundation project aims to bring DNS-style trust to AI agents. [online] 25 June. Available at: https://www.infoworld.com/article/4189361/new-linux-foundation-project-aims-to-bring-dns-style-trust-to-ai-agents.html [Accessed 15 July 2026].
Linux Foundation (2026 ). Linux Foundation Announces Intent to Launch Agent Name Service to Establish Trusted Identity Infrastructure for AI Agents. [online] 23 June. Available at: https://www.linuxfoundation.org/press/linux-foundation-announces-intent-to-launch-agent-name-service-to-establish-trusted-identity-infrastructure-for-ai-agents [Accessed 15 July 2026].
Merced, A. (2026 ). The State of Agentic AI Standards in 2026: MCP, A2A, WebMCP, OSI, and the Protocol Stack Taking Shape. DEV Community. [online] 7 July. Available at: https://dev.to/alexmercedcoder/the-state-of-agentic-ai-standards-in-2026-mcp-a2a-webmcp-osi-and-the-protocol-stack-taking-3o2l [Accessed 12 July 2026].
Mohammed, A.R., Joy, A. and Challa, K. (2026 ). From Visibility Gaps to Agentic Operations: The AI Supply Chain Tower Story. Oracle AI & Data Science Blog. [online] 12 June. Available at: https://blogs.oracle.com/ai-and-datascience/from-visibility-gaps-to-agentic-operations-the-ai-supply-chain-tower-story [Accessed 12 July 2026].
Persistent Systems (2026 ). Agentic Enterprise in 2026: Why AI Needs Operating System, Not Another Pilot. [online] 22 June. Available at: https://www.persistent.com/blogs/agentic-enterprise-in-2026-why-ai-needs-operating-system-not-another-pilot/ [Accessed 12 July 2026].
Unanimous Technologies (2026 ). Rise of AI Agents in Modern Applications in 2026. [online] 19 June. Available at: https://unanimoustech.com/rise-of-ai-agents-in-modern-applications/ [Accessed 12 July 2026].
Uplift (2026 ). The State of Automation and AI-Agent Adoption 2026. [online] Available at: https://getuplift.ai/resources/state-of-automation-ai-agent-adoption-2026 [Accessed 12 July 2026].
