Core concepts

Two complete use cases

AI-first software delivery covers understanding a request and repository, accepted specification, coding-agent work, tests, independent review, evaluation, approval, evidence, and release decision.

Agentic product runtime covers models, tools, context, workflows, human decisions, policy, telemetry, and outcomes inside the delivered software.

llm-core owns the stable contracts between those use cases. AIFSD composes those contracts into products and host plans. A qualified runtime adapter executes them through its native framework.

Ownership

flowchart TB
  Delivery["AIFSD SDK / CLI / delivery application"]
  Kernel["llm-core contracts · authority · conformance · evidence"]
  Runtime["Qualified runtime adapter"]
  Native["LangGraph / PydanticAI / Strands / other native runtime"]

  Delivery --> Kernel
  Kernel --> Runtime
  Runtime --> Native

The kernel does not own an agent loop, workflow engine, conversation executor, scheduler, or durable worker fleet.

Explicit extension boundaries

• AgentDefinition describes portable agent intent.
• AgentRunner is the port implemented by a runtime integration.
• CompiledSpecification<T> binds a projected value to reviewed authority.
• native values remain native behind explicit integration-owned surfaces.
• specification adapters declare a closed matrix of all five exact operations as supported, unsupported, or not-applicable; every entry binds the declaration's recognised source contract exactly, and supported entries carry immutable fixtures.
• normalized events and evidence enable comparison without claiming identical behavior.

See API by subpath for the package surface.