The agent platform landscape of 2026

If you survey what “agent platform” means in 2026, the field has quietly split into five categories that barely talk to each other. Most arguments about which one is winning are really arguments about which problem the speaker thinks agents are for.

The first category is hyperscaler-managed runtimes. AWS Bedrock AgentCore, Google Vertex Agent Engine, Microsoft Foundry Agent Service, Cloudflare Agents. These exist because large enterprises will not run unmanaged Python processes that talk to LLMs, and the hyperscalers know it. The pitch is IAM, VPC, audit, compliance. The price is per-vCPU-hour with a long bill of separately-metered components. The lock-in is real but legible.

The second is open-source frameworks. LangGraph, CrewAI, AutoGen, Mastra, LlamaIndex, Pydantic AI, Letta, Smolagents, the Claude and OpenAI Agent SDKs. You run them yourself. They differ on language (Python or TypeScript), on orchestration primitive (graph, roles, handoffs, sub-agents), and on what they take seriously about state. Persistence is usually an afterthought bolted on with a checkpointer.

The third is workflow automation that learned to spell “agent.” n8n, Zapier, Make, Activepieces, Pipedream, Dify, Langflow. These are deterministic recipe engines with LLM nodes welded in. Every one of them shipped MCP support in 2025 or 2026, which makes them more useful as integration substrate than as agent platforms. The boundary is sharp: if the work can be expressed as a DAG known in advance, a workflow tool will do it cheaper and more reliably than an agent. If it cannot, no number of LLM nodes will save you.

The fourth is durable execution engines. Temporal, Restate, Inngest, Hatchet, Trigger.dev, Dapr Agents. The thesis is that agent reliability is not a prompt-engineering problem, it is a journaling problem. LLM calls are non-deterministic and expensive. Naive retries re-bill and produce different outputs. Crashes are inevitable. The fix is to journal every non-deterministic step and replay from cache. Temporal pioneered the pattern. Everyone else is converging on it. Restate’s “virtual objects” and Hatchet’s Postgres-only durable tasks are the most interesting reads on the same idea.

The fifth, and the one that actually shapes most enterprise budgets, is the vertical in-product copilot. Sierra for customer service. Harvey for legal. Hippocratic for healthcare. Cooper for Kaseya. Sidekick for ConnectWise. Robin for Atera. N-zo for N-able. These are not platforms at all. They are a model, a prompt, a curated tool set, and a SKU. They sell the outcome, not the substrate.

Five categories. The interesting thing is what is missing.

There is no winner yet in the MSP layer — the category-creation move where someone wraps a heterogeneous stack of RMM, PSA, M365, and documentation tools in a single agentic orchestrator that the MSP owns. Rewst is closest, but Rewst is a workflow tool, not an agent. The in-product copilots only see their own vendor’s data. The hyperscalers will not ship MSP-specific tools. The open-source frameworks ship the substrate but not the integrations.

The thing that makes 2026 different from 2025 is the MCP protocol becoming the actual USB-C of this world. Anthropic crossed 10,000 public MCP servers in March. Every major RMM now ships one, or has been wrapped by community ones. Pipedream’s managed-OAuth MCP gateway exposes 10,000 tools across 3,000 apps. The integration tax that used to gate every agent project has collapsed.

What hasn’t collapsed is the orchestration question. The category boundaries above exist because each platform answered “what is the unit of work” differently. A Temporal workflow is not a Cloudflare durable object is not a LangGraph node is not an n8n step is not a Sierra agent. You cannot mix and match. You pick one and commit, or you build your own and pay the orchestration tax yourself.

I think the honest read of the landscape is this: the substrate question is mostly answered. The integration question is mostly answered. The two unanswered questions are durable replay of LLM calls without re-billing the model, and vertical orchestration layers in industries where no incumbent has shipped one yet. Those are the two places where new products can still define a category in 2026.

Everything else is execution.