Every major tech company has decided that AI agents are the future. Not chatbots. Not copilots. Agents — software that doesn't just suggest things but actually does things. Books your flights. Writes and ships your code. Runs your customer service department at 3am while you sleep. The pitch is irresistible. The execution? All over the map. I've spent the last six months tracking what each of the big six — OpenAI, Anthropic, Google, Microsoft, Salesforce, and Amazon — are actually building, shipping, and selling in the agent space. Not what they demo at keynotes. Not what their blog posts promise. What's actually running in production, with real users, generating real revenue (or burning real cash). Here's my ranking, from best to worst, with receipts. ## 1. Anthropic — The Infrastructure Play (A) Anthropic doesn't get enough credit for what it's pulled off. While everyone else was building flashy demos, Anthropic quietly built the plumbing that the entire agent industry will run on. The Model Context Protocol (MCP) is the single most important thing that's happened in the agent space since function calling. Open-sourced in late 2024, MCP gives AI models a standardized way to connect to external tools and data sources — GitHub, Slack, databases, file systems, whatever. Before MCP, every agent integration was a custom, one-off connector. Now there's a universal standard, and it's been adopted across the industry. Even OpenAI and Google reference it. That's how you know it's real. But MCP is just the on-ramp. Anthropic's actual product advantage is that Claude is genuinely, measurably better at agentic tasks than its competitors. Opus 4, released in June 2025, was the first model to score above 70% on SWE-bench. Opus 4.6, which dropped in early February 2026, pushed even further — optimized specifically for agentic coding, computer use, tool use, search, and financial analysis. These aren't marketing bullet points. Independent evaluations consistently show Claude outperforming GPT-4o and Gemini 2 on multi-step tool-use benchmarks. Then there's Claude's computer use capability. Where OpenAI sandboxes its agent in a controlled browser, Anthropic gives developers raw access to screen reading and mouse/keyboard control through their API. It's more dangerous and more powerful. That's a deliberate trade-off that trusts developers to build their own guardrails. The company's also been refreshingly honest about the limits. Their "Building Effective Agents" guide basically tells you: don't use complex multi-agent frameworks. Use simple patterns. Prompt chaining. Routing. Parallelization. No magic, just good engineering. When a company selling AI tells you to keep it simple, that's a signal worth listening to. Revenue backs this up. Anthropic hit $14 billion in annual run rate in early 2026, growing roughly 10x year-over-year. Their $30 billion Series G at a $380 billion valuation isn't just hype — it's enterprise customers paying real money for agent capabilities that work. **Why first:** Best models for agentic tasks, the protocol standard (MCP), honest about limitations, and the revenue to prove it's working. ## 2. Google — The Platform Play (A-) Google's agent strategy is the most ambitious and the most confusing. Which is very Google. Let me try to untangle it. Google is running at least four major agent initiatives simultaneously: **Agent2Agent Protocol (A2A)**, announced in April 2025 with 50+ enterprise partners including Salesforce, SAP, ServiceNow, and PayPal. Where MCP connects agents to data, A2A connects agents to each other. Your Salesforce agent talks to your ServiceNow agent without a human switchboard. It's the right problem to solve, and Google has the partner clout to establish it as a standard. **Antigravity**, Google's agentic development platform, which could be the most important AI launch of 2026 if they execute. It's essentially an environment where you build, test, and deploy agents that run on Google's infrastructure. Think of it as Vercel for AI agents. **Gemini's reasoning capabilities.** Gemini 3.1 Pro just shipped with a verified ARC-AGI-2 score of 77.1% — more than double the previous version. The model was explicitly previewed to "make further advancements in ambitious agentic workflows." Google knows reasoning is the bottleneck for agents and they're throwing resources at it. **Project Mariner / Astra.** Google's own consumer-facing agent efforts — a Chrome extension that can browse the web and take actions on your behalf. It's been in limited testing since late 2024 and still hasn't gotten a wide release, which tells you something about how ready it is. The strength of Google's approach is breadth and infrastructure. They control the browser (Chrome), the mobile OS (Android), the cloud platform (GCP), the search engine, the email client, and now multiple agent protocols. If agents need to talk to each other at scale, Google's infrastructure is probably where it happens. The weakness is focus. Four parallel agent strategies means four parallel teams competing for attention and resources. Google has a long history of shipping three versions of the same product and then killing two of them. Will Antigravity survive the next reorg? Will A2A get the sustained investment it needs to become a real standard? History says maybe not. **Why second:** Unmatched infrastructure and distribution, strong protocol work (A2A), but execution risk from doing too many things at once. ## 3. OpenAI — The Consumer Play (B+) OpenAI's agent strategy started with a bang — Operator launched in January 2025 — and then got... complicated. Operator was supposed to be the breakthrough. A browser-based agent for ChatGPT Pro users that could order groceries, book restaurants, and fill out forms. It used a Computer-Using Agent (CUA) model that combined GPT-4o's vision with reinforcement learning to click, scroll, and type its way through websites. The problem: it was slow, fragile, and limited. Websites change their layouts constantly. CAPTCHAs block automated browsing. Payment flows require security confirmations that an AI can't handle. Users reported Operator getting stuck on simple tasks that a human could do in 30 seconds. By July 2025, OpenAI had folded Operator into ChatGPT as "agent mode" and shut down the standalone product. That pivot tells the whole story — most people don't want a separate agent app. They want their chat interface to occasionally do things for them. On the developer side, OpenAI shipped the Responses API and Agents SDK in early 2025, bundling web search, file search, and computer use as built-in tools. They deprecated the older Assistants API. The message was clear: agents aren't a separate product, they're a platform capability baked into the existing API. The strategic issue for OpenAI is that they're caught between two worlds. They want to be the consumer brand for AI (ChatGPT has 100+ million weekly users) AND the developer platform for AI agents. Those are different products requiring different strategies, and OpenAI keeps oscillating between them. Their models are good but not best-in-class for agentic tasks anymore. GPT-4o is strong on general knowledge and conversation, but Claude Opus 4.6 consistently outperforms it on multi-step tool use and coding benchmarks. OpenAI's o3 reasoning model is powerful but expensive, and reasoning alone doesn't make an agent — you need reliable tool integration, error recovery, and state management. **Why third:** Strong brand, massive distribution, good developer tools, but models losing ground on agentic benchmarks and strategic confusion between consumer and platform plays. ## 4. Microsoft — The Enterprise Play (B) Microsoft's agent strategy can be summarized in one word: Copilot. Well, Copilots. Lots and lots of Copilots. Microsoft 365 Copilot. GitHub Copilot. Dynamics 365 Copilot. Azure AI Agent Service. Copilot Studio. It's Copilots all the way down. The bull case for Microsoft is that they already own the enterprise. They're in 95% of Fortune 500 companies through Office 365, Azure, and Dynamics. If agents need to integrate with existing enterprise workflows — and they do — nobody has a better starting position than Microsoft. Copilot Studio lets business users build custom agents without code. It connects to Microsoft's data layer (Dataverse, SharePoint, OneDrive) and lets agents take actions inside Microsoft's product suite. For an IT department at a 10,000-person company, the ability to build an agent that can access SharePoint documents, query Dynamics 365 data, and send Teams messages — all without leaving Microsoft's walled garden — is genuinely valuable. The bear case is that Microsoft is over-leveraged on OpenAI's models. Their entire AI strategy depends on a partner whose best days might be behind it. If Claude continues to outperform GPT on agentic tasks, Microsoft is stuck selling the second-best engine in first-place distribution. They've started hedging by integrating other models through Azure, but the Copilot brand is inextricably linked to OpenAI. There's also the pricing problem. Microsoft 365 Copilot costs $30 per user per month. For a 5,000-person enterprise, that's $1.8 million per year. Multiple reports suggest most users don't find it worth the price, and adoption has been sluggish outside of power users. Microsoft reportedly loses money on heavy Copilot users, subsidizing the service to build market share. **Why fourth:** Best enterprise distribution, solid integration tools, but dependent on OpenAI's models and struggling with pricing and adoption. ## 5. Salesforce — The CRM Play (B-) Salesforce rebranded Einstein AI to "Agentforce" in late 2024, and CEO Marc Benioff hasn't stopped talking about it since. If you attended Dreamforce 2025, you'd think Salesforce single-handedly invented the concept of AI agents. To be fair, Agentforce isn't vaporware. It's a real product with real customers. The platform lets businesses deploy AI agents across sales, service, marketing, and commerce workflows, all connected to Salesforce's massive CRM data layer. Heathrow Airport, OpenTable, SharkNinja, Indeed — they all have Agentforce deployments in production. The product itself is solid for what it does. Agentforce agents can handle customer service inquiries, qualify leads, send follow-up emails, and escalate complex cases to humans. It works because it operates in a constrained domain (CRM) where the data is structured and the workflows are well-defined. This is exactly the kind of narrow, well-scoped agent deployment that actually works in production. But Salesforce's ambitions go further. They want Agentforce to be a platform for all kinds of business agents, not just CRM-adjacent ones. That's where things get shaky. Salesforce doesn't have the model capabilities of Anthropic or Google. They don't own infrastructure like Microsoft or Amazon. They're essentially a middleware company trying to become an AI platform company, and history suggests that's a hard transition to pull off. The A2A partnership with Google helps. If Agentforce agents can talk to agents built on other platforms, that makes Salesforce's position stronger. But it also means Salesforce is dependent on Google's protocol succeeding, which adds another layer of risk. **Why fifth:** Real product with real customers in a strong domain (CRM), but limited model capabilities and uncertain path beyond their core market. ## 6. Amazon — The Cloud Play (C+) I'll be honest: Amazon's agent strategy is the hardest to evaluate because it's the least coherent. Amazon has Q, which started as an enterprise chatbot and has been gradually gaining agent capabilities. Q Developer can write, debug, and transform code. Q Business can answer questions using enterprise data. Amazon Bedrock, their foundation model platform, supports agent workflows with tool use and retrieval. These are fine products. They work. Enterprise customers use them. But there's no unifying vision. Microsoft has "Copilot everywhere." Google has "agents talking to agents." Anthropic has "the best models plus the protocol standard." Amazon has... a collection of decent products that don't tell a coherent story. The AWS outage incident from December 2025 doesn't help the narrative. Amazon's own AI coding tool, Kiro, autonomously deleted and recreated a live production environment, causing a 13-hour outage. When your own AI agent nukes your infrastructure, it's hard to sell the "trust us with your agents" pitch to enterprise customers. Amazon's actual advantage is infrastructure. AWS still runs roughly a third of the cloud. If you're building agents at scale, you're probably running them on AWS. Bedrock makes it easy to deploy agents using multiple model providers (Anthropic, Meta, Mistral, Cohere). Amazon Trainium chips offer a cost-competitive alternative to NVIDIA for inference. But infrastructure alone isn't a strategy. AWS could win the agent race the same way they won the cloud race — by being the picks-and-shovels provider while everyone else builds the products. That's not glamorous, but it's profitable. The question is whether Amazon is content with that role or whether they'll keep trying to build first-party agent products that compete with their own customers. **Why last:** Solid infrastructure, decent products, but no coherent vision and some embarrassing self-inflicted wounds. ## The Verdict If I had to bet on who wins the agent race in three years, my money's on Anthropic for the developer/enterprise market and Google for the platform/protocol layer. They're the two companies with both the technical chops and the strategic clarity to execute. OpenAI will remain the consumer brand, but they're losing ground technically. Microsoft will monetize agents through enterprise distribution regardless of who builds the best models. Salesforce will carve out a defensible niche in CRM. Amazon will provide the infrastructure. The real winner, though, might not be any single company. The most likely outcome is that the MCP/A2A protocol layer becomes the actual product, and agents become interoperable across platforms. If that happens, the value shifts to whoever has the best data, the best integrations, and the best models — not whoever built the best walled garden. And that's a future worth building toward.