Module agent 

Agent runtime — AgentRunner, Orchestrator, workflow agents, guardrails, and supporting types.

Modules

audit

Agent audit trail — structured records of LLM calls, tool invocations, and completions.

batch

Batch executor for running the same agent on multiple tasks with controlled concurrency.

blackboard

Shared blackboard for inter-agent state exchange in squad-based orchestration.

cache

LLM response caching for deterministic replay and cost reduction.

context

Context management strategy for long-running agent sessions.

dag

DAG (Directed Acyclic Graph) workflow agent.

debate

Debate workflow agent.

evaluator

Evaluator-Optimizer workflow agent.

events

Agent lifecycle events emitted during execution via the OnEvent callback.

guardrail

Guardrail trait and GuardAction — base types for all guardrail implementations.

guardrails

Built-in guardrail implementations — LLM judge, secret scanner, PII detector, content fence, action budget, behavioral monitor, injection classifier, tool policy, and composition helpers.

handoff

Agent handoff runner.

instructions

Dynamic instruction loading from .heartbit/INSTRUCTIONS.md files.

mixture

Mixture-of-Agents (MoA) workflow agent.

observability

Observability mode — controls tracing verbosity and metric emission.

orchestrator

Multi-agent orchestrator for parallel and sequential sub-agent delegation.

permission

Runtime permission rules for tool usage — deny, allow, and learned-via-HITL decisions.

prompts

Built-in system prompt fragments for multi-agent collaboration.

pruner

Session context pruner — trims old messages before LLM calls to control context size.

routing

Dynamic complexity-based agent routing.

tenant_tracker

Per-tenant in-flight token tracker with Arc-owning RAII reservations.

token_estimator

Lightweight token estimation for context budget management.

tool_filter

Dynamic tool filtering based on query classification.

voting

Majority-voting workflow agent.

workflow

Deterministic workflow agent primitives.

Structs

AgentOutput

Output of a completed agent run.

AgentRunner

Runs an agent loop: LLM call → tool execution → repeat until done.

AgentRunnerBuilder

Builder for AgentRunner.

Type Aliases

OnInput

Callback for interactive mode. Called when the agent needs more user input (i.e., the LLM returned text without tool calls). Returns Some(message) to continue the conversation, or None to end the session.