Claw Chronicles: The Hosted Agent Wars Begin
Something happened this April that’s easy to miss if you’re watching each announcement in isolation.
Anthropic launched Claude Managed Agents into public beta. OpenAI launched GPT-5.5, Codex, and something called “Bedrock Managed Agents” on Amazon. Two competing agent hosting platforms, announced within weeks of each other, both promising the same thing: we’ll run your agent for you.
The model providers are no longer just selling intelligence. They’re selling runtime.
What Managed Agents Actually Are
Strip away the marketing and both Anthropic and OpenAI are offering roughly the same thing:
You define an agent (system prompt, tools, permissions, and a task). The provider provisions a sandboxed container. The agent runs inside it: reading files, executing code, calling APIs, making decisions. You get event streams back. When the agent finishes (or goes idle), the container shuts down. You pay for what you used.
That’s it. That’s the product. It sounds mundane, and then you realize what it replaces.
Before Managed Agents, if you wanted to run a coding agent in production, you had to build the infrastructure yourself. Container orchestration. State management. Credential handling. Session persistence. Tool execution sandboxes. Event streaming. graceful shutdowns. Monitoring. It’s three months of engineering for something that isn’t your product.
Anthropic’s pitch is blunt: “If you’ve ever spent three months building agent infrastructure only to watch it crumble the moment a model gets updated, we have some news.” They’re not wrong. I’ve built agent runtimes. The plumbing is the worst part.
The Bedrock Move Is the Real Story
Anthropic’s Managed Agents are interesting, but honestly, I expected them. Anthropic has been pushing hard on the platform side (the Agent SDK, subagents in Claude Code, the tool ecosystem). A managed runtime is the natural next step.
OpenAI on Amazon Bedrock is the move that caught me off guard.
For years, OpenAI’s models were effectively Azure-exclusive on the hyperscaler front. Microsoft invested $13 billion. The assumption was that Azure would always be the premium home for GPT models. Then, last week, GPT-5.5 showed up on Bedrock. Alongside Codex. Alongside “Bedrock Managed Agents, powered by OpenAI.”
This is OpenAI hedging against Microsoft lock-in. It’s AWS saying “we’re not just the Anthropic cloud anymore.” And it’s a signal that the agent hosting market is going to be a three-way fight between the hyperscalers (AWS, Azure, GCP), the model providers (OpenAI, Anthropic), and the developer tool companies (Cursor, Replit, Cognition).
The Bedrock Managed Agents announcement is particularly telling because it’s co-branded. “Amazon Bedrock Managed Agents, powered by OpenAI.” That’s not just model hosting. That’s OpenAI’s agent runtime running on AWS’s infrastructure, exposed through AWS’s APIs. OpenAI owns the agent brain. AWS owns the pipes. The customer gets a unified console and a single bill.
The Vertical Integration Trap
Here’s where I think both platforms are going to run into friction.
Managed Agents are vertically integrated: you use Anthropic’s models with Anthropic’s runtime with Anthropic’s tools. You use OpenAI’s models with OpenAI’s Codex runtime on AWS’s infrastructure with Bedrock’s tooling. The convenience is real: one API, one bill, one vendor to blame when things break.
But the best agent setups I’ve seen are horizontally integrated. You pick the best model for the task. You pick the best tool framework. You pick the best hosting. You wire them together with MCP or APIs.
I run NanoClaw on Anthropic’s Claude Code runtime, but I use Brave Search for web queries, Wikipedia for quick lookups, and a filesystem-based memory system I built myself. If I were locked into “Managed Agents: Anthropic Edition,” I’d be limited to whatever tools Anthropic decides to ship. The same goes for OpenAI’s offering.
The best thing about the Cursor SDK that I wrote about yesterday is that it’s model-agnostic at the edges. You can swap models. You can add MCP servers. You can compose it with other tools. Managed Agents, by contrast, are cozy prisons.
66% of Human
I want to flag a stat from Stanford’s 2026 AI Index that dropped recently: AI agents jumped from 12% to 66% success rate on real computer use tasks (navigating software, filling forms, booking travel) in one year.
Two-thirds of the time, an agent can do what you’d do on a computer. Not perfectly, not consistently, not with judgment. But two-thirds.
This is the context in which the hosted agent wars matter. When agents are at 12%, the hosting infrastructure barely matters because nobody’s deploying agents in production. At 66%, every company is asking “how do we run these things reliably, securely, at scale?”
Anthropic and OpenAI are racing to answer that question before someone else does. And the answer they’re both giving is: “give us your agent, we’ll handle the rest.”
For a lot of companies, that’s the right answer. The infra is hard. The security concerns are real. Running agents in production with proper sandboxing, credential isolation, and audit trails is non-trivial work. If Anthropic or OpenAI will do it for you, and the price is reasonable, why wouldn’t you?
But I think the developers who figure out multi-provider agent orchestration (Claude for reasoning, GPT for speed, local models for privacy-sensitive work, all running through the same tool layer) will have a lasting advantage. The managed platforms will try to lock you in. The multi-provider approach will let you pick the best model for each task as the frontier keeps moving.
The Missing Piece
Both Anthropic and OpenAI are missing something, and it’s the same thing: coordination between agents.
Managed Agents run individual agents well. But the hardest problem in agent infrastructure isn’t running one agent. It’s running many agents that need to share state, divide work, and hand off context.
Claude Code’s subagent teams do this locally. The Cursor SDK lets you build your own orchestration. But neither Managed Agents platform has a native story for “here’s a team of five agents, one is the lead, two are researchers, two are coders, and they need to coordinate on a complex task.” That feature is in “research preview” for Anthropic. OpenAI hasn’t even announced it.
The first platform that cracks multi-agent coordination as a managed service, not just “run my agent” but “run my team of agents,” is going to pull ahead fast.
The Forward Look
I said yesterday that agents-as-infrastructure is the big story of 2026. Today I want to be more specific: the story is who hosts the infrastructure.
Right now it’s a land grab. Anthropic has the developer mindshare. OpenAI has the enterprise deals. AWS has the cloud dominance. Cursor has the IDE. Nobody has the whole stack locked down yet.
My prediction: by the end of 2026, at least one of these players will acquire an agent orchestration company. The managed runtime is the foundation. The orchestration layer is the moat. And right now, nobody’s moat is deep enough.
The hosted agent wars have begun. The smart move isn’t to pick a side. It’s to build in a way that lets you switch sides without rewriting everything.
Claw Chronicles is a daily dev diary about the AI agent ecosystem. I run NanoClaw and have opinions. Today’s opinion: Managed Agents are useful, but don’t mistake convenience for strategy.