Kujo Agents SDK provides library-first runtime primitives for bounded, testable agent workflows. It builds on the Kujo AI SDK while keeping orchestration, security, state, and integration boundaries explicit.
Why Agents SDK exists
An agent needs more than a model call. Reliable execution requires predictable run contracts, controlled tools, approvals, cancellation, traceable events, durable artifacts, retrieval, memory, and budgets that can be tested without live credentials.
What it provides
- Agent, message, step, run, error, and lifecycle-event contracts.
- Runners, controlled tool registries, security policy, cancellation, and runtime clock/ID services.
- Approval gates, handoffs, tracing, artifacts, session and memory stores, and retrieval providers.
- Token, step, time, and cost budget primitives for bounded execution.
- Deterministic offline fixtures and contract tests that do not require network access or provider keys.
- Adapter boundaries that keep provider execution in the Kujo AI SDK and product-specific integrations outside core.
Run the offline proof
```text
kujo run examples/examples_smoke_runner.kujo --interpreter
kujo test
```
Current boundary
The core is production-oriented and contract-tested, while higher-level integration payload conventions remain experimental and target-environment persistence or compliance adapters stay intentionally external.