A product of ANH Digital Platform · For Cargill researchers & dairy consultants
A runtime for the work that doesn’t fit in a chat window.
Maestro treats the language model as a reasoning kernel inside a distributed runtime, not as the application. Research takes weeks, not seconds. Data arrives continuously. Costs compound. The chatbot-shaped stack handles none of that well. Maestro is the answer — built on three primitives the current stack lacks.
Three primitives
Everything else — MCP, vector stores, graph orchestrators — remains. But they move to the edges. The center of Maestro is these three.
Continuations
The unit of work is not a model call. It is a resumable, migratable reasoning process that can sleep for weeks and wake on a signal.
Read more → 02Context Field
A continuously-projected, ranked surface over every memory source — not a database the agent queries, but a field it reads.
Read more → 03Policy Kernel
A small, deterministic, sub-10ms decision service that routes model tiers, enforces budgets, and gates tools. No LLMs on the critical path.
Read more →Where we’re breaking ground
The 2026 stack converged on graphs over MCP. That stack handles chat-shaped work. It does not handle research-shaped work — jobs that run for weeks, branch and merge, sleep and wake, accumulate evidence, and have to stay inside a budget. Maestro re-bases the architecture so it does.
Resources
The source artifacts behind this site. The architecture doc is the canonical reference; the YAML plan is what Claude Code consumes.
The future of agentic work is continuations over streams, not graphs over RPCs.
That is the architectural bet. The rest of this site documents how Maestro implements it, what we have built so far, and what is coming next. The architecture is a living document — this site rebuilds whenever it changes.