The agent runtime landscape in 2026
Six months ago, no one had shipped an agent runtime. Now five vendors are fighting for the category. A no-BS comparison of Bedrock AgentCore, Claude Managed Agents, Microsoft Agent Framework, LangGraph Cloud, and Appstrate, with the axes that actually decide which one you want.
🗺️ Five platforms, one category, very different bets
A year ago, if you asked "where should I run this agent in production?", the answer was a shrug. You picked a framework, wired it to your web server, hoped for the best. There was no category.
By March 2026, there are five credible platforms that call themselves agent runtimes and mean roughly the same thing by it. They solve the same problem (execute an agent as a sandboxed, stateful, multi-tenant process) and bet different parts of the stack.
This post is the comparison I wanted when we started building Appstrate. It is not a neutral piece: we make one of the five. But the table reflects the real trade-offs, and we'll be explicit about where each one wins.
📊 The table
| Bedrock AgentCore | Claude Managed Agents | Microsoft Agent Framework | LangGraph Cloud | Appstrate | |
|---|---|---|---|---|---|
| Vendor | AWS | Anthropic | Microsoft / Azure | LangChain | Open-source |
| Self-hostable | ❌ | ❌ | Partially (via Azure) | Partially (via LangGraph OSS) | ✅ |
| License | Proprietary | Proprietary | Proprietary | MIT (OSS core) + closed cloud | Apache 2.0 |
| BYOM | Bedrock models + bring-your-own-endpoint | Claude only (+ limited partners) | Azure OpenAI-first + open weights | Any (it's your graph) | Any (provider-neutral) |
| Sandboxing | Firecracker microVM | Proprietary sandbox | Azure Container Apps | Your infra | Docker + sidecar (Firecracker roadmap) |
| Multi-tenant primitives | AWS accounts / IAM | Anthropic workspaces | Entra / Azure AD | Limited | Orgs + apps + end-users + impersonation |
| Credential broker | IAM-mediated | Partial | Key Vault-backed | DIY | Sidecar, per-provider allowlists |
| Scheduling | EventBridge | Limited | Azure Logic Apps / Functions | DIY | BullMQ-based, first-class |
| Webhooks | EventBridge | Limited | Event Grid | DIY | First-class, Standard Webhooks spec |
| Headless API | ✅ | ✅ | ✅ | ✅ | ✅ (191 endpoints) |
| Versioned API | Date-based | Per-feature | Date-based | Semver on SDK | Date-based (Appstrate-Version) |
| Best for | AWS-native teams | Pure-Claude teams | Azure-native enterprises | Teams with complex graphs | Everyone else |
🧠 One paragraph per player
Each of these platforms is a serious piece of engineering. What distinguishes them is the bet they made about who their user is.
AWS Bedrock AgentCore: the cloud-native bet
Bedrock AgentCore is the most infrastructure-rich of the five. Firecracker microVMs, IAM-mediated tool access, EventBridge for webhooks, CloudWatch for logs, KMS for keys. If your whole stack is already on AWS, AgentCore slots in and inherits the operational patterns your team already knows. The flip side is the one you'd expect: you're running on AWS, paying AWS prices, locked into AWS primitives, and every cross-region story is an AWS cross-region story. Great for AWS shops. A non-starter for data-residency edge cases and for teams whose security review treats cloud vendor lock-in as a risk.
Anthropic Claude Managed Agents: the model-native bet
Claude Managed is the smoothest path from a prompt to a running agent, as long as the prompt is for Claude. The runtime is tuned to Claude's tool-calling, Claude's thinking, Claude's computer-use, Claude's caching. For a team that has decided Claude is the right model and wants to ship, Claude Managed eliminates a lot of wiring. For a team that wants to evaluate another model seriously, or hedge against pricing changes, or run a local fallback, you're on the wrong platform by construction.
Microsoft Agent Framework: the enterprise-native bet
The most interesting thing about MSAF is that it is partially self-hostable through the OSS components, which makes it the only managed runtime with a coherent story for Azure Stack and sovereign clouds. If your procurement already runs through Microsoft Enterprise Agreements, there is a strong path for you here. The trade is that the full experience is Azure-first, Entra-first, and the best tooling assumes you're a .NET shop. If you are, it's a very clean option. If you aren't, you're second-class.
LangGraph Cloud: the graph-native bet
LangGraph Cloud is the only platform of the five whose primary abstraction is the graph of the agent, not the agent as a black box. If your agents have complex, explicit state machines (multi-agent orchestration, human-in-the-loop, long-running with checkpoints), LangGraph's model is the closest to how you already think. The trade is that you are responsible for the infrastructure around the graph: deployment, multi-tenancy, scheduling, audit. LangGraph Cloud manages the orchestrator; it doesn't manage the platform around it. Excellent if you want a well-designed execution engine. Not a platform replacement.
Appstrate: the open-source bet
Appstrate is the only one of the five that is Apache-2.0, self-hostable by default, and provider-neutral on models. We believe the agent runtime category will play out the way the cloud category did: there will be three or four proprietary winners with strong managed products, and one or two open-source alternatives that handle the long tail of teams who need control of the infrastructure: regulated industries, SaaS builders embedding agents, cost-sensitive workloads, sovereign deployments, teams burned once by a vendor rug-pull. We're building for that long tail. The bet is that the category gets big enough that the long tail is a real business, the way Linux and Postgres became real businesses.
🎯 The axes that actually decide
If you're evaluating the five, a pros-and-cons list is less useful than a decision on a handful of axes. Here's the short version.
Model strategy
If you've already committed to one model family and don't see that changing, pick the runtime that's tuned to it: Claude Managed for Claude, AgentCore for Bedrock families, MSAF for Azure OpenAI. If you haven't committed, or you plan to keep options open, model-neutrality matters more than model-specific tuning. That's LangGraph or Appstrate.
Where the data has to live
If your data cannot leave a specific region, certain clouds, or your own perimeter, the question isn't "which runtime". It's "which runtime will let me run in that perimeter at all". On that axis, Appstrate and the self-hosted parts of LangGraph / MSAF are the remaining set. The others are out by design.
Multi-tenant model
If you are running agents on behalf of your users (not just your team), the multi-tenant primitives matter a lot. AWS IAM and Azure Entra are powerful but designed for enterprise identity, not end-user embedding. Stripe-style end-users with impersonation, API keys scoped to applications, and webhooks designed for B2B SaaS is a specific product requirement that not all five solve the same way.
Ops tolerance
If your team has zero ops bandwidth, self-hosting is a losing trade, regardless of the OSS arguments. Pick a managed runtime, pay the tax, ship. If your team can run Postgres / Redis / Docker on Kubernetes or a fleet of VMs without flinching, self-hosting unlocks the price, sovereignty, and flexibility wins that managed platforms can't match.
Community and portability
The closed platforms have excellent vendor support. What they don't have is a second implementation you can switch to without a rewrite. Apache-2.0 code with a clean internal API is a different kind of insurance. It doesn't protect you from outages, but it protects you from the vendor disappearing, changing license, or moving the product upmarket.
🧭 A short decision guide
Because the table above is a lot to hold in your head, here's the shorter version:
- You're AWS-native, Claude-native, or Azure-native, and happy about it: pick the vendor runtime that matches. The integration tax is real and the vendor platforms pay it for you.
- You're building agents inside a regulated or sovereign deployment: Appstrate or self-hosted LangGraph. Probably Appstrate if you want the multi-tenant primitives too.
- You're building a SaaS that embeds agents for your users, branded, with your auth: Appstrate. The end-user + impersonation + webhooks model is designed for exactly this shape.
- You're running genuinely complex graphs of agents with human-in-the-loop checkpoints: LangGraph. The graph abstraction is the differentiator.
- You're running a lot of agents and the managed bill is eating your margin: self-host something. Appstrate is the cheapest trade between infra you have to run and platform you have to build.
🔓 Why we bet on open-source
An honest close, since we're one of the five.
Every mature infrastructure category in the past twenty years ended up with the same shape. Proprietary leaders with excellent products (VMware, Oracle, Splunk, Datadog). Open-source alternatives that captured the long tail and eventually pulled down the proprietary margins (KVM, Postgres, OpenTelemetry, Grafana). The open-source alternatives didn't win because they were better. They won because they gave teams a control plane they owned in a world where the stakes of not owning it kept going up.
Agent runtimes are the next category on that trajectory. The stakes are high: these systems hold your users' credentials, run on their data, take actions with real consequences. Teams are already feeling the need for a control plane they can open, read, fork, and run where they want. That is the long tail we are building for.
None of this implies the proprietary platforms will lose. Bedrock, Claude Managed, MSAF will be the default for a large chunk of the market, just like Oracle is still the default for a large chunk of databases. What it implies is that there should be a credible open-source option for the other chunk. That's the category Appstrate is trying to fill.
📚 Related
Other posts
- What is an agent runtime platform?: the category definition behind the table above.
- Why personal agents leak credentials: the sandboxing axis, in detail.
- Your agents, your infrastructure: the self-hosting argument.
- Claude Code for teams: the multi-tenant axis, in detail.
Appstrate docs
- Compare: the docs version of this comparison, with more rows.
- Concepts: the Appstrate architecture in full.
- Quickstart: five minutes to your first run.