Module runtime 

Harness runtime facade.

AgentHarness is the single composed runtime that every level of the recursion runs inside: a sub-agent, a subgraph node, a REPL session, or a model-authored blueprint all execute on the same harness — the same model and tool registries, middleware stack, and RunPolicy. That shared, re-entrant runtime is what makes “agents calling agents” and self-authored workflows recurse on one consistent set of capabilities rather than spinning up disjoint engines.

Owns the high-level AgentHarness builder and the RunPolicy bundle that wires registries, middleware, and run policy into a single ergonomic entry point. The agent loop driven by this facade lives in the sibling crate::harness::agent_loop module.

Layout

• [types] holds the public type definitions (RunPolicy and AgentHarness).
• This file holds the builder, registration, and accessor methods.
• test.rs holds focused tests for construction and registration.

Structs

AgentHarness

High-level facade that composes model selection, tool execution, middleware, and run policy behind one builder and runs the default agent loop.

PayloadCapture

Controls whether the agent loop captures model and tool payloads (prompt/completion text, tool arguments/results) onto the AgentEvent::ModelCompleted and AgentEvent::ToolCompleted events it emits.

RunPolicy

Enums

InvalidArgsPolicy

How the agent loop reacts when the model calls a registered tool with arguments that fail schema validation.

UnknownToolPolicy

Declarative, run-scoped policy shared by every invocation of an AgentHarness.