“Agentic AI technology is the software stack that lets an enterprise hand a system an outcome to achieve rather than a script to execute”  © Dr. Harish Kotadia, 2026 — the model, memory, tools, identity, guardrails, and observability that turn a goal into governed, auditable action. Instructions in, results out — that was IT. Intent in, outcomes out — that is agentic AI.

A Practitioner’s Long Definition of Agentic AI

Agentic AI is the next enterprise workload abstraction. In my pillar definition, What is Agentic AI?, I define it as a governed, goal-driven software layer in which LLM-powered agents — equipped with memory, tools, and orchestration protocols — autonomously plan and execute multi-step business processes across your cloud, data, and application estate, the way containers once abstracted servers and pipelines once abstracted data movement. Where a traditional application runs the steps you code, an agent plans the steps itself, calls tools, reads and writes to your systems, checks its own work, maintains memory across steps, operates under guardrails, and escalates to a human at defined boundaries it should not cross alone.

That is the line separating genuine agentic AI from what Gartner calls “agent washing” — chatbots and RPA scripts relabeled. Genuine agentic AI requires autonomous decision-making, multi-step reasoning, and dynamic error handling. Instructions in, results out was IT; intent in, outcomes out is agentic AI.

I have spent more than two decades architecting enterprise systems for Fortune 100 clients, and in my current work building multi-agent origination systems on Anthropic Claude for regulated loan origination, I have learned that the model is rarely the hard part. The hard part is the stack around it: memory, tools, identity, guardrails, observability, and the human-in-the-loop review that decides when an agent may act.

How Big Is This Market, Really?

Agentic AI is growing faster than almost any enterprise technology in memory while deploying into production more slowly than almost any other. Gartner forecasts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. In Gartner’s best-case scenario, “agentic AI could drive approximately 30% of enterprise application software revenue by 2035, surpassing $450 billion, up from 2% in 2025.”

But adoption is not production. McKinsey’s “The State of AI in 2025” (published November 5, 2025; 1,993 respondents across 105 nations) found that 88% of organizations regularly use AI in at least one function, yet “Twenty-three percent of respondents report their organizations are scaling an agentic AI system somewhere in their enterprises,” and “In any given business function, no more than 10 percent of respondents say their organizations are scaling AI agents.” Gartner (press release, June 25, 2025) expects more than 40% of agentic AI projects to be canceled by the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls. That gap between intent and production is exactly where the technology choices in this post matter most.

 

The Agentic AI Technology Stack: Layer by Layer

The agentic AI stack has ten layers in my map. Each answers a different engineering question. I define each layer answer-first, then list the technologies in demand.

Layer 1: Foundation and Frontier Models

Foundation models are the reasoning core of an agent — the component that plans, decides which tool to call, and drafts the output. In 2026 the frontier compressed: no single model wins every task, so most serious enterprises route across several.

• Anthropic Claude — Claude Opus 4.8 (shipped May 28, 2026) leads real-world coding and agentic tool use, topping SWE-bench Verified at 88.6% and ranking #1 on the Artificial Analysis Intelligence Index. It is my default for regulated, multi-step work where reasoning quality and tool reliability matter more than raw context size. Priced at $5 / $25 per million input/output tokens.
• OpenAI GPT-5.5 — released April 23, 2026 and now ChatGPT’s default; strongest on terminal-based coding agents and long-context retrieval. Available through Azure AI Foundry and the OpenAI API. Priced at $5 / $30 per million tokens — the most expensive output of the big three.
• Google Gemini 3.1 Pro — largest context window and strongest multimodal and novel-reasoning results (77.1% on ARC-AGI-2), and the cheapest of the three flagships at $2 / $12 per million tokens.
• Open-weight models — Meta Llama, Mistral (strong for European data residency), and DeepSeek for self-hosted, privacy-sensitive, or fine-tuned workloads. These anchor the “run it in our own VPC” pattern.

The pattern that wins is not “use Claude” or “use GPT.” It is a routing strategy: send each task to the cheapest model that clears the quality bar, and reserve the frontier model for the reasoning-heavy steps.

Layer 2: Cloud AI Platforms

Cloud AI platforms are the managed runtimes where enterprises actually host and govern agents, inheriting identity, storage, compliance, and observability from the cloud they already run. Three hyperscaler platforms dominate.

There is no single best. Vertex leads on model breadth, Bedrock on AWS-native compliance and guardrails, Azure AI Foundry on the Microsoft and OpenAI ecosystem. IBM watsonx is the fourth option, purpose-built for regulated industries needing governance and auditability. My advice is unchanged from my earlier writing: start on the cloud you already run, because a second cloud triples networking, observability, and compliance overhead.

Layer 3: Agent Frameworks and Orchestration

Agent frameworks are the libraries that give you the coordination primitives — how agents plan, hold state, call tools, hand off work, and recover from failure — so you are not rebuilding message passing and checkpointing from scratch. This layer consolidated sharply in the first half of 2026.

• LangGraph (LangChain) — the default for stateful production workflows in regulated industries. Graph-based state machines, durable checkpointing, human-in-the-loop interrupts, time-travel debugging. Roughly 400 companies now run agents on LangGraph Platform, including Klarna, Uber, and JPMorgan. My pick when auditability and deterministic control matter.
• CrewAI — fastest path to a role-based multi-agent prototype; roughly 46,000–48,000 GitHub stars in early-to-mid 2026, 100,000+ certified developers, with native MCP and A2A support.
• Microsoft Agent Framework — reached 1.0 GA on April 3, 2026, unifying AutoGen and Semantic Kernel into one .NET and Python SDK with native MCP and A2A. The default for Azure/.NET teams.
• OpenAI Agents SDK — minimal, handoff-based, lowest friction for GPT-centric agents with sandboxed tools.
• Anthropic Claude Agent SDK — deepest MCP integration and lifecycle hooks; strongest for coding agents with OS access.
• Google ADK — multi-language (Python, Java, TypeScript, Go), native A2A Agent Cards; roughly 8 million cumulative downloads within its first year.
• LlamaIndex — event-driven orchestration for document-heavy, data-intensive pipelines.

AutoGen and Semantic Kernel are now maintenance-only, superseded by Microsoft Agent Framework; new projects should not start on AutoGen.

Layer 4: Interoperability Protocols and Standards

Agent interoperability protocols are the standard wires that let agents reach tools and reach each other without a bespoke integration for every pairing. Two protocols now define the layer.

• Model Context Protocol (MCP) — Anthropic’s open standard (November 2024) for connecting an agent to tools, data, and APIs — “USB-C for AI agents.” Per Anthropic’s December 9, 2025 announcement donating MCP to the Linux Foundation’s Agentic AI Foundation, MCP had reached 97 million-plus monthly SDK downloads with roughly 10,000 active servers — a figure that surpassed 110 million monthly downloads by April 2026 — with OpenAI, Google, Microsoft, and AWS all shipping support.
• Agent2Agent (A2A) — Google’s protocol for agent-to-agent collaboration, donated to the Linux Foundation (June 2025) and reaching v1.0 in April 2026, backed by more than 150 organizations including Salesforce, SAP, and ServiceNow.

MCP connects an agent to tools; A2A connects agents to each other. They are complementary, not competing. In my own work, MCP-based tool access with human-in-the-loop review is the backbone of the origination systems I build. In 2026, protocol literacy is a core enterprise-architect skill.

Layer 5: Enterprise Platform-Embedded Agents

Platform-embedded agents are the agents your SaaS vendors ship inside the systems of record you already run, so agents act on live data with inherited permissions and audit. This is where the biggest production numbers live.

• Salesforce Agentforce — the fastest-growing product in Salesforce history. Per Salesforce’s Q1 FY27 earnings release (May 27, 2026), Agentforce ARR reached $1.2 billion, up 205% year over year, with 3.8 billion Agentic Work Units delivered across Agentforce and Slack. Built on Data Cloud with the Atlas reasoning engine and the Einstein Trust Layer.
• Microsoft Copilot Studio — more than 230,000 organizations (including 90% of the Fortune 500) have used it, and customers created over 1 million custom agents in a single quarter across SharePoint and Copilot Studio (Microsoft, Build 2025).
• ServiceNow AI Agents / Now Assist — Now Assist annual contract value crossed $600 million in 2025, more than doubling year over year, and reached $750 million entering Q1 2026; ServiceNow raised its full-year AI ACV target to $1.5 billion. Built on the Now Platform with native CMDB access.
• IBM watsonx Orchestrate — reached GA at Think 2026 (May 2026) with 150-plus enterprise connectors, repositioned as an agentic control plane; IBM’s generative AI book of business surpassed $12.5 billion since inception (IBM Q4 2025 earnings, January 28, 2026). Positioned as the orchestration and governance layer for regulated enterprises.
• SAP Joule, Oracle AI Agent Studio, Workday Illuminate Agents, Adobe Experience Platform Agent Orchestrator — every major suite vendor now embeds agents, reshaping SaaS pricing toward usage- and outcome-based models.

Layer 6: Memory, Retrieval, and the Knowledge Layer

The knowledge layer is what gives an agent memory and grounding — the vector store and RAG tooling that let it carry context across steps and answer from your data instead of starting cold every turn. Several vector databases anchor production in 2026.

• Pinecone — managed leader, scales to billions of vectors with low operational burden.
• pgvector — the Postgres extension and pragmatic default; production-grade at Supabase, Neon, and Instacart scale.
• Weaviate — native hybrid search (vector + keyword), strong for diverse content.
• Milvus and Vespa — the billion-vector-plus tier.
• Qdrant — Rust-based open-source speed and filtering;
• Chroma — the prototyping developer-experience leader.

Hybrid search (vector + keyword + metadata filters) has become non-optional for agent memory and RAG, because agents need exact matches for IDs and version numbers alongside semantic matching. My standard first move for any enterprise is a RAG pipeline over governed internal documents as the shared knowledge layer for all agents.

Layer 7: Governance, Identity, Security, and Observability

This layer answers three questions Okta frames well for the agentic enterprise: where are my agents, what can they access, and what are they allowed to do. It is the layer most enterprises under-build.

• Non-human identity (NHI) — Okta’s “Okta for AI Agents” (GA April 30, 2026) treats each agent as a first-class identity with its own lifecycle, an agent gateway, and vaulted credentials. Per Gravitee’s State of AI Agent Security 2026 report (February 3, 2026), cited by Okta: “88% of organizations report confirmed or suspected AI agent security incidents. Only 22% have identities tied to their agents.” CyberArk and SailPoint address the same NHI gap.
• Observability and evaluation — LangSmith (deepest LangChain/LangGraph integration), Langfuse (open-source, self-hostable), and Arize Phoenix (OpenTelemetry-native, ML-grade drift detection).
• AI gateways and guardrails — LiteLLM (open-source routing to 100+ providers), Portkey (managed gateway with guardrails and PII redaction), Kong AI Gateway, and NVIDIA NeMo Guardrails for policy-as-code. Gateways increasingly govern MCP and A2A traffic, not just model calls.

Security and risk concerns are now among the top barriers to scaling agentic AI. This is the layer that moves an organization from Piloted to Governed on my roadmap — and the one most programs treat as a follow-on step rather than foundational work.

Layer 8: Coding and Developer Agents

Coding agents are the most mature, highest-adoption category of agentic AI — the one where “agent” already means real autonomous work, not autocomplete. AI coding tool usage among engineering teams jumped to 73% daily use in 2026 from 41% the year before (Pragmatic Engineer survey, February 2026, 906 developers).

• Claude Code — terminal-native, strongest for hard multi-file reasoning; reached an estimated $2.5 billion annualized run rate by February 2026, up from $500 million in September 2025. Preferred by 46% of senior developers versus 9% for Copilot in JetBrains’ 2026 survey.
• GitHub Copilot — the enterprise default with 20M+ users, deepest GitHub integration and compliance; strongest at large enterprises (56% adoption at 10,000+ employee firms).
• Cursor — the leading AI-native IDE; Anysphere crossed $2 billion in annualized revenue by February 2026, with roughly 60% of revenue now from enterprise contracts.
• OpenAI Codex — named a Leader in Gartner’s April 2026 Magic Quadrant for enterprise AI coding agents; surpassed 5 million weekly active users as of June 2026 (OpenAI).
• Devin (Cognition) — the most autonomous “AI software engineer,” now paired with Devin Desktop (formerly Windsurf).

The dominant enterprise pattern is a two-layer stack: Copilot deployed broadly as the baseline, plus Claude Code adopted by senior engineers for high-leverage agentic work.

Layer 9: Infrastructure and Runtime

This layer is the secure sandbox where agent-generated code and tool calls actually run without risking the host. As agents move from producing text to taking actions, a guardrail failure means a bad action — data deleted, money moved — so isolation matters.

• E2B — open-source, Firecracker microVM isolation with ~150ms cold starts; E2B’s own materials claim 94% of Fortune 100 companies have signed up (company-reported).
• Daytona — container-based, sub-90ms creation, persistent workspaces (production codebase went closed-source June 2026).
• Modal, Blaxel, Runloop, Northflank — managed sandboxes with GPU and bring-your-own-cloud options.

Defense in depth — compute isolation plus filesystem, network, and resource limits — is the correct architecture; no single layer suffices.

Layer 10: The Ecosystem View — How the Layers Fit Together

An enterprise agentic AI system is these layers assembled, not any one product. In practice a frontier model (Layer 1) runs inside a cloud platform runtime (Layer 2), orchestrated by a framework (Layer 3), reaching tools and other agents over MCP and A2A (Layer 4), grounded by a vector/RAG knowledge layer (Layer 6), often acting through platform-embedded agents in your systems of record (Layer 5), executing untrusted steps in a sandbox (Layer 9), authored partly by coding agents (Layer 8), and wrapped end-to-end by an identity, guardrail, and observability governance plane (Layer 7).

Most large programs run both platform and framework: platform for breadth and procurement-friendly governance, framework for depth and control. NVIDIA underscored the consolidation at GTC 2026 (March 2026), launching the NVIDIA Agent Toolkit with 17 software partners including Salesforce, SAP, ServiceNow, and Adobe. The ecosystem is converging on one architecture: acting agents on trusted enterprise data, wrapped in governance.

This is exactly the maturity my Five-Stage Agentic AI Roadmap describes. The technologies in Layers 6 and 7 — memory, identity, guardrails, observability — are what move an organization from Piloted to Governed. Most enterprise deployments today cluster at Governed and thin out sharply toward Assured; none yet credibly evidence Autonomous.

In-Demand Skills and Roles for Enterprise Agentic AI

The scarce skill in 2026 is no longer configuring a platform — it is orchestrating, grounding, and governing agents on top of it. McKinsey (August 2025) reported that “agentic AI job postings grew 985% last year,” and Stanford’s 2026 AI Index separately found agentic-AI postings up more than 280% year on year, reaching roughly 90,000 US postings.

• Agentic AI Architect — designs the multi-agent blueprint: which agents exist, how they communicate, what memory they share, how failures are handled. Glassdoor (2026) lists average U.S. AI Architect pay at $189,788 per year, with top earners up to $336,915. This is the fastest-growing enterprise role as agents move to production.
• Agentic AI Engineer — builds the agent loops: tool calling, sub-agent orchestration, memory, and evaluation harnesses. Framework fluency (LangGraph, CrewAI, the vendor SDKs) plus eval design is the hardest skill to fake.
• AI Governance / Responsible AI Lead — owns accountability, EU AI Act and sector compliance, and agentic-AI risk. Forrester’s 2026 predictions state that the complexities of navigating the patchwork of legislation “will cause 60% of the Fortune 100 to hire or appoint a designated head of AI governance,” noting Sony, Bank of America, and UBS have already done so.
• AgentOps / AI Operations Lead — keeps production agents running: eval pipelines, cost monitoring, incident response. A natural pivot for SRE and MLOps engineers.
• Agentic Data Specialist — bridges data engineering with agent needs: RAG architecture, vector databases, data governance.
• AI Product Manager (Agents) — defines what an agent should and should not do, where humans review, and the eval criteria.

The through-line: the highest-paid professionals are T-shaped — broad foundational knowledge with deep expertise in one layer. Protocol literacy (MCP, A2A), evaluation rigor, and governance are the differentiators that command a premium.

Conclusion

The agentic AI landscape looks sprawling, but it resolves into ten layers and one principle: instructions in, results out was IT; intent in, outcomes out is agentic AI. The winners in 2026 are not the organizations with the biggest model budget. They are the ones treating the unglamorous layers — memory, identity, guardrails, observability — as foundational work rather than a follow-on step, and placing themselves honestly on the roadmap by their weakest production agent, not their best pilot.

Pick the layers your use case actually needs, start on the cloud you already run, and build the governance plane before you scale. That is how a pilot becomes production.

If you are mapping your own agentic AI stack, tell me which layer you are stuck on — that is usually where the real architecture work begins.

---

© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026

Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.

He is the author of Intent In, Outcomes Out: A Practitioner’s Field Guide to Agentic AI for the Enterprise, available on Amazon at: https://mybook.to/AgenticAI

He writes the Agentic AI Architect newsletter on LinkedIn: https://www.linkedin.com/newsletters/agentic-ai-architect-7477724166675140608/

What is Agentic AI? Agentic AI is a governed, goal-driven software layer in which LLM-powered agents autonomously plan and execute multi-step business processes. Instructions in, results out, was IT. Intent in, outcomes out, is agentic AI. Read the full definition at: https://agenticaiarch.com/what-is-agentic-ai/

The Agentic AI Roadmap is a five-stage maturity model — Prompted, Piloted, Governed, Assured, Autonomous — measuring how reliably and accountably an organization turns intent into outcomes. Explore the roadmap at:
https://agenticaiarch.com/agentic-ai-roadmap/

Follow him on LinkedIn at www.linkedin.com/in/hkotadia and at www.agenticaiarch.com