Agent Runtime
Bounded autonomy with an explicit blast radius.
The execution environment where agents plan, call tools, and act, under policy limits they cannot exceed.
Why it exists
Autonomy is only acceptable when its limits are explicit. An agent that can open a ticket and an agent that can restart a production cluster need the same reasoning engine and radically different permissions, approval paths, and rollback guarantees.
How it works
Planning and decomposition
Work is broken into steps with declared preconditions and expected effects, so the plan can be inspected before any of it runs.
Typed tool invocation
Agents act only through registered tools with typed schemas, argument validation, and per-tool rate and scope limits. There is no generic shell.
Policy gates
Each step is evaluated against policy: allowed environments, value thresholds, change windows, and whether human approval is required before execution.
Rollback and idempotency
Actions declare their compensating operation. Retries are idempotent by construction so a partial failure never produces a duplicate payment or a double remediation.
What it gives you
- Inspectable plans with declared preconditions and effects
- Typed, scoped tool registry, no arbitrary execution
- Policy gates for environment, value, and change windows
- Human-in-the-loop approval on configurable thresholds
- Idempotent retries with declared compensating actions