Scaling Managed Agents:
Decoupling the Brain from the Model

A running theme in agent infrastructure is that harnesses encode assumptions — about what models cannot do, about where code runs, and about which model is doing the thinking. Those assumptions go stale fast.

Anthropic described this well in their recent engineering blog post on Claude Managed Agents. They found that workarounds built for Sonnet were dead weight on Opus. Context resets, retry logic, premature-completion guards — all tuned for one model's limitations, all wrong for the next.

We hit the same wall, but from a different angle. We were not just swapping model versions. We were swapping model providers. A planning agent running Opus. A code agent running Sonnet. A review agent running GPT for a second opinion. A local agent running Ollama for sensitive data that cannot leave the machine.

We built Lattice Runtime to solve both problems: a managed agent infrastructure that is opinionated about the interfaces — brain, session, hands — but not about which model is the brain or where the hands live.

The same old problem, one layer up

Anthropic framed their challenge as an old problem in computing: how to design a system for “programs as yet unthought of.” Operating systems solved this by virtualizing hardware into abstractions general enough for programs that did not exist yet.

We agree with the pattern. We disagree with where they drew the boundary.

Anthropic virtualized everything except the model. Their brain interface assumes Claude. Their session lives on Anthropic's cloud. Their sandbox runs in Anthropic's containers. They decoupled the components — brain, session, hands — but coupled them all to a single provider.

Lattice Runtime virtualizes one layer further: the model itself becomes a pluggable interface.

// The brain doesn't know what model is behind this.
response := brain.Generate(ctx, GenerateRequest{
    Messages: session.GetEvents(lastCheckpoint),
    Tools:    runtime.AvailableTools(),
})

// Could be Claude, GPT, Gemini, or Ollama.
// Config change, not code change.

The brain calls generate(messages, tools) → response — it does not know if Claude, GPT, Gemini, or a local Ollama instance is behind it. The session lives wherever you run it. The abstraction outlasts the provider.

Do not adopt a vendor

Anthropic described the “pets vs cattle” problem: when everything lives in one container, that container becomes a pet you cannot afford to lose.

We had the same problem, but the pet was not a container — it was a vendor.

When your managed agent infrastructure is Claude-only, your entire agent system becomes a pet tethered to one provider's API, pricing, uptime, and context window.