
Modular Agent Runtime (MAR)
The problem
Running an agent's turn reliably needs a lot of plumbing: agent lookup, context assembly, model and tool access, policy, budgets, retries, cancellation, delegation, hooks, state, and memory gateways. That plumbing is easy to duplicate badly across projects, or to tangle with business logic.
The core layer
MAR defines the mechanics of a run as typed contracts around a kernel — a
RunRequest in, a RunResult out. Composition resolves agents, models, tools, policy, context, state, and memory; control enforces capabilities, budgets, and retries; extensibility adds hooks and delegation; observability emits events and traces. MAR never owns the business meaning of the application built on top of it.The implementation
It's a real Python package with a quickstart example that wires a model provider, a tool, an agent definition, and a FastAPI boundary without adding domain code into MAR itself. Both the core package and the example's own integration tests are verified with pytest.
0 likes · 5 views · Sep 18, 2026
A typed, modular runtime for reliably executing AI agent turns without mixing infrastructure and application logic.
Timeline
Aug 1, 2026 - Aug 31, 2026