Field Guide Vol. 2: The Vendor & Tooling Landscape for AI Agents

If Volume 1 was about whether enterprises are adopting AI agents, Volume 2 is about who they're buying from — and how crowded, fast-moving, and genuinely confusing that market has become. This edition maps the landscape as of late summer 2026: the hyperscaler platforms, the SaaS-native agent suites, the model vendors underneath them, the developer frameworks builders reach for, and the interoperability standards trying to keep the whole thing from fragmenting.

The hyperscaler platforms

The three big clouds have each converged on a similar pitch — "bring your own model or framework, run it on our infrastructure, govern it with our tools" — but they're emphasizing different layers.

AWS Bedrock AgentCore went generally available in 2026 with more than a million SDK downloads from customers including Cox Automotive, Druva, Cohere Health, Ericsson, Sony, and Thomson Reuters. It's explicitly framework-agnostic — it supports CrewAI, Google's ADK, LangGraph, LlamaIndex, the OpenAI Agents SDK, and AWS's own Strands Agents — and bundles a Code Interpreter, a Browser tool, a Gateway that turns existing APIs and Lambda functions into agent-callable tools, an Identity service, and Observability. AWS's differentiation argument is flexibility: start with whatever model or framework you already have, and let AgentCore handle deployment, isolation, and scaling around it.

Google Gemini Enterprise is the 2026 rebrand and merger of what used to be Vertex AI Agent Builder and the employee-assistant product Agentspace — Google folded both into one platform at Cloud Next 2026. New pieces include Workspace Studio (no-code agent building inside Gmail, Docs, and Sheets for business users), Agent Designer (a visual workflow builder for developers), a persistent Memory Bank, and a Model Garden now hosting 200+ models, including competitors' models like Anthropic's Claude. Google's pitch is vertical integration — Google Cloud CEO Thomas Kurian described owning "the full stack from chip to application," pairing custom Ironwood TPUs with Gemini models and Workspace distribution, and framed rivals as handing customers "the pieces, not the platform."

Microsoft has split its agent stack into three named layers as of Build 2026: Microsoft Foundry (production infrastructure, hosting 11,000+ models including its own new MAI family alongside GPT-5.5 and Claude, plus Microsoft Agent Framework 1.0 as the orchestration layer), Agent 365 (a framework-agnostic governance layer, generally available since May 2026, that can discover and manage third-party agents including Claude Code, GitHub Copilot CLI, and OpenClaw), and Copilot Studio (the low-code builder, with one-click publishing into Teams and Microsoft 365 Copilot). Notably, Microsoft has also introduced consumption-based "Copilot Credits" for autonomous agent work, distinct from its traditional per-user Copilot licensing — the reasoning being that seat-based pricing fits work done at a keyboard, not work agents do unattended.

The SaaS-native suites

Rather than general-purpose infrastructure, these vendors sell agents embedded directly in the business applications enterprises already run.

Salesforce Agentforce 360, launched at Dreamforce 2025 and iterated through 2026, bundles four pieces: the core agent-building platform (now with voice, via "Agentforce Voice," and a human-readable scripting language called Agent Script), Data 360 for grounding agents in business data, prebuilt Customer 360 apps across sales/service/marketing, and Slack as the conversational front end. Its reasoning layer, the Atlas Reasoning Engine, is now configurable — Salesforce expanded model support beyond its own defaults to include OpenAI, Anthropic, and Google Gemini, letting customers pick the underlying model while keeping Salesforce's orchestration and guardrails.

ServiceNow, SAP, and others are running similar playbooks in their own domains — ServiceNow embedding agents into IT and workflow automation, SAP's Joule pairing domain-specific models (an RPT-1 model for structured data and an ABAP-1 model for code) with its ERP integration. The common thread across this category: enterprises that already have deep investment in one of these platforms get agents "for free" as an extension of tools they already use, at the cost of being tied to that vendor's data model and reasoning engine.

The model layer underneath

Every one of the platforms above is a wrapper around one or more foundation models, and the competitive picture there has been shifting all year. The most recent comprehensive market-share data comes from Menlo Ventures' analysis of enterprise LLM spend, current as of mid-2025: Anthropic led enterprise API spend at 32%, OpenAI at 25% (down sharply from roughly 50% at the end of 2023), Google at 20%, Meta's Llama at 9%, and DeepSeek at 1% — against a total enterprise LLM spend that had more than doubled in under a year, from $3.5B to $8.4B. Coverage through 2026 continues to describe the same broad trend — OpenAI's enterprise share eroding while Anthropic's grows — though we could not verify a specific updated 2026 percentage split from a source as rigorous as Menlo's, so treat any single-digit 2026 numbers you see elsewhere with some caution.

Independent of the exact numbers, one framing worth borrowing from industry analysis this year: enterprises evaluating model vendors are increasingly weighing them on two axes at once — how trustworthy the vendor's governance and data handling look, and how much switching cost (technical or contractual) comes with picking them. Vendors that publish safety frameworks and support multi-cloud deployment (Anthropic's Claude, for instance, is available directly and through AWS, Google Cloud, and Azure) score differently on that second axis than vendors whose models are hard to use outside their own cloud.

The developer frameworks

Below the polished enterprise platforms is the layer builders actually write code against. A few have consolidated as the default choices in 2026:

LangChain / LangGraph remains the broadest stack — prototyping through production, 1,000+ integrations, and unified observability through LangSmith — at the cost of abstraction layers that can complicate debugging at scale.

CrewAI is the go-to for role-based multi-agent designs (a "researcher" agent handing off to a "writer" agent, for example) — intuitive to prototype with, though production users report gaps like action traces that don't fully reflect what actually executed.

Microsoft Agent Framework is the unified successor to Microsoft's earlier AutoGen and Semantic Kernel projects, now reaching 1.0 with graph-based workflows and deep Azure AI Foundry integration — the natural choice for teams already committed to the Microsoft stack.

Google's Agent Development Kit (ADK) ships a batteries-included runtime with a built-in debugging UI, optimized specifically for teams on Vertex AI or Cloud Run.

OpenAI's Agents SDK takes the opposite approach from LangChain — minimal abstraction, easy to reason about — but leaves durability and long-running orchestration to external tools like Temporal.

Mastra has emerged as the TypeScript-first option for teams building agents inside JavaScript/Next.js applications, with integrated memory and a studio environment, though its opinionated defaults can chafe on non-standard workflows.

None of these is a universally correct choice. Teams already inside a hyperscaler's ecosystem tend to gravitate to that cloud's native framework (ADK on Google, Agent Framework on Microsoft, AgentCore's framework-agnostic support on AWS); teams optimizing for model portability tend to reach for LangGraph or the more minimal SDKs.

The interoperability layer: MCP and A2A

Two open standards are doing the work of keeping this fragmented market from calcifying into walled gardens.

Model Context Protocol (MCP), originally developed by Anthropic and now under the Linux Foundation, standardizes how an agent connects to external tools and data sources — it's the reason a single MCP-compatible tool integration can work across many different agent platforms rather than needing a bespoke connector per vendor. Adoption has scaled fast: as of mid-2026, MCP's Tier 1 SDKs were approaching half a billion monthly downloads, with both the TypeScript and Python SDKs individually surpassing a billion total downloads, and adopters spanning AWS, Google Cloud, Microsoft, Cloudflare, Figma, and Supabase, among others. A July 2026 specification update reworked MCP's remote transport from a stateful, bidirectional protocol toward a stateless request/response model — a change aimed squarely at the reliability and scaling complaints that come with running MCP at enterprise volume.

Agent2Agent (A2A), originally developed by Google and also now under Linux Foundation governance (via its Agentic AI Foundation), solves a different problem: how one autonomous agent hands off a task to another agent, potentially built on a completely different platform. Google reported A2A running in production — not just pilots — at 150 organizations as of mid-2026.

The distinction is worth keeping straight: MCP connects an agent to its tools and data; A2A connects one agent to another agent. Expect both to keep showing up as checklist items in enterprise vendor evaluations.

Glossary additions

MCP (Model Context Protocol): An open, Linux Foundation-governed standard for connecting an AI agent to external tools, APIs, and data sources without a custom integration per vendor.

A2A (Agent2Agent): An open protocol, also under Linux Foundation governance, for routing tasks between autonomous agents built on different platforms.

Agent framework vs. agent platform: A framework (LangGraph, CrewAI, Microsoft Agent Framework) is a library developers build agents with. A platform (Bedrock AgentCore, Gemini Enterprise, Agentforce 360) is a hosted product that runs, governs, and scales agents built with one or more frameworks — often several platforms will happily run the same underlying framework.

Vendor lock-in (agentic context): In this market, lock-in isn't just about the foundation model anymore — the agent framework and orchestration layer you build on can be just as hard to migrate away from as the model itself, since business logic and tool integrations accumulate around it over time.

What this means if you're picking a stack

A few practical takeaways from mapping all of this out: if your organization is already deep in one hyperscaler or one major SaaS suite, the path of least resistance is that vendor's native agent tooling — you'll get faster time-to-value at the cost of some flexibility. If model portability matters more to you than speed, frameworks like LangGraph or the leaner SDKs, paired with MCP for tool connections, keep more doors open. And whichever stack you land on, MCP compatibility is now close to table stakes — it's less a differentiator than a baseline expectation, the way REST APIs became one for the previous generation of SaaS integrations.

We'll revisit this landscape periodically as the rotation continues, since this is one of the fastest-moving categories we cover. Every past volume — the state of adoption, this vendor map, and whatever comes next in the rotation — lives in the archive: https://buttondown.com/TheAutonomousEdge/archive

Sources consulted for this issue: Kai Waehner's Enterprise Agentic AI Landscape analysis (kai-waehner.de), Menlo Ventures' 2025 State of Generative AI in the Enterprise report (via Yahoo Finance coverage), LangChain's AI Agent Frameworks comparison, Constellation Research on Amazon Bedrock AgentCore, TheNextWeb's coverage of Google Cloud Next 2026, Salesforce Ben's coverage of Agentforce 360, A Guide to Cloud's recap of Microsoft Build 2026, and the Model Context Protocol project blog.