R / 02 / Active programme
Agents as bounded systems, not unstructured loops.
We study the runtime responsibilities beneath capable agents: process isolation, scheduling, identity, budgets, signals, memory, checkpointing, evaluation, and an auditable event stream.
Research question
What must an agent runtime guarantee before autonomous work can become durable, governable, and recoverable?
Workstreams
The programme is worked as a system.
Runtime semantics
Define agents as supervised processes with lifecycle, scheduling, budgets, signals, and explicit termination behaviour.
State & durability
Study checkpointing, replay, memory, event history, and crash recovery across long-running units of work.
Authority & protocols
Bound identity, tools, IPC, MCP, A2A, secrets, and delegation so capability does not silently become permission.
Evaluation & observability
Connect task quality to runtime traces, policy decisions, resource use, human intervention, and reproducible failure.
Evidence gate
What would move the claim forward?
Research maturity advances through named evidence, not through the passage of time or the confidence of the presentation.
- 01Deterministic event history
- 02Crash and resume behaviour
- 03Budget and permission enforcement
- 04Task and runtime evaluation
Current boundary
The limit is part of the result.
Namzu is available as an open-source TypeScript agent kernel and SDK. Individual subsystems and integrations retain their own documented maturity and release status.