Anthropic published their Managed Agents architecture. I've been running a similar pattern for months from Dome, my local-first knowledge management app. Here's how they compare and why both approaches are valid.
Anthropic just published their architecture for Managed Agents: decoupling the “brain” (the LLM and its harness) from the “hands” (sandboxes, tools, execution environments). Their core insight is that harnesses encode assumptions about what models cannot do, and those assumptions go stale as models improve.
I have been running a version of this pattern for months. Not on Anthropic’s infrastructure. From Dome, my local-first knowledge management app.
What Dome’s agent harness actually does
Dome has native MCP (Model Context Protocol) support. This means AI agents can interact directly with notes, resources, and external tools through a unified interface that abstracts exactly the execute(name, input) pattern Anthropic describes.
The pipeline I built works like this: I write a note in Dome, tag it with [publicar], and a manual automation triggers three MCP-connected agents in sequence. The first reads the draft from Dome and generates bilingual MDX with complete Astro frontmatter, adapting the English version without literal translation. The second writes both files to the local repository via Filesystem MCP. The third runs git add, commit and push via Git MCP, which triggers CI/CD. Two minutes later the post is live in Spanish and English. Dome then updates the note title to mark it published.
The repository as CMS
The git repository is the CMS. The commit history is the publication log. The agents are the editors.
Where Dome’s approach differs from Anthropic’s Managed Agents
Anthropic’s approach is excellent infrastructure. But it comes with tradeoffs that matter depending on your context.
On cost and API dependency: Managed Agents run on Anthropic’s platform, which means every agent action is an API call. At scale, that compounds. Dome runs locally. LLM inference can be Ollama, LM Studio, or any GGUF-compatible model running on your own hardware. Zero API cost per agent loop.
On LLM flexibility: Anthropic’s harness is built around Claude. Dome’s harness is model-agnostic. The same MCP tools, the same workflow, work with Claude, Mistral, Llama, Gemma, or whatever runs locally. You are not locked to one provider’s intelligence curve.
On data control: In Managed Agents, your session data, context, and agent memory live in Anthropic’s infrastructure. In Dome, everything stays local by default. Notes, resources, agent outputs, none of it leaves your machine unless you explicitly push it somewhere.
On harness customization: Anthropic has opinions about the interfaces (session, harness, sandbox) but not about the implementations. Dome gives you the same separation: MCP defines the interface, and you wire whatever tools you want behind it. The difference is you control the wiring.
The shared architectural insight
Both Anthropic’s Managed Agents and Dome’s harness converge on the same principle: decouple context from execution from tools. The abstractions outlast the implementations.
Anthropic built this as managed cloud infrastructure for enterprise scale. I built it in a desktop app because I wanted to publish blog posts without a CMS.
The patterns are the same. The tradeoffs are different. Both are valid depending on what you are optimizing for.
Many brains, many hands, whether those hands are Anthropic’s sandboxes or your local filesystem and a git remote.
