Skip to main content
Version: latest

agrag.agents

Agentic layer: one task-scoped evidence agent over the graph.

build_agent is imported lazily because it needs the optional agents extra. Importing this package does not require that extra.

Modules:

  • build – Build the single task-scoped evidence agent.
  • errors – Errors raised by the agent layer.
  • harness – Process-global DeepAgents harness profile registration.
  • ledger – Citation ledger: assigns and tracks stable keys for one agent run.
  • middleware – Middleware for composing configured chat models and capturing task events.
  • model – Translate LLMClientConfig into the matching LangChain chat model.
  • prompts – The single default agent prompt.
  • result – Typed result returned after a task-scoped agent run.
  • settings – Env-backed LLM and loop config for the agent layer.
  • tools – Tools bound to one task run: evidence, navigation, task and decision tools.
  • tracing – Per-run OpenInference tracing for agent runs.
  • verification – The structured verdict of the fixed-input verifier.

Classes:

  • AgentLLMSettings – LLM client config for the agent's own reasoning turns.
  • AgentMissingExtraError – Agent tracing needs a package extra that is not installed.
  • AgentRunResult – Outcome, answer, evidence references and conversation for one run.
  • AgentSettings – Configuration for the agent loop itself.
  • Ledger – Assigns and tracks stable citation keys for one agent run.

Functions:

  • build_agent – Build one agent bound to trusted document and memory scopes.