Agent and Model Development

An agent is a bounded package-defined capability. A model is a replaceable inference capability. A workflow is the durable process that invokes agents, deterministic steps, review gates and tools. They have distinct identities and lifecycles.

Implement the invocation contract

Use the platform execution contract for actor, tenant, purpose, schema versions, authorized context references, budgets and trace identity. Return schema-valid outputs, evidence references, uncertainty or abstention, model/tool provenance and a bounded completion state. A package does not invent new authority from retrieved content.

Model capabilities

Bind a task to a declared ModelPackage capability instead of embedding one provider or model name in business logic. Qualification identifies exact model, runtime, quantization, adapter, hardware and task/language profile. A replacement must satisfy the consuming package's schema, quality, latency and policy requirements.

Tools and effects

Declare tool schemas, requested scopes, timeout and retry behavior. Use the gateway for policy enforcement and audit. External writes are reviewable intents with approval, idempotency and compensation; agents do not hold unrestricted credentials or write around the gateway.

Framework adapters

Native SDK implementations and qualified framework adapters map to the same canonical invocation, event, tool, output and audit contracts. A framework session does not replace the durable workflow or environment lock. See runtime technology profiles for the compatibility baseline.

Evidence and evaluation

Reference source versions and the precise supporting fragments for material assertions. Test unsupported claims, conflicting evidence, prompt injection, unauthorized sources, ambiguous identities and low-coverage cases. Domain evaluation fixtures belong to the package; the platform supplies execution, recording and admission controls.

Concrete implementations

The agent ownership index links all supplied agents to their canonical package reference. Those manuals contain agent-specific triggers, inputs, outputs, tools, limits and metrics. Reuse a declared capability rather than creating a second conflicting copy of its agent.