Skip to content

API Reference

Auto-generated reference documentation for all public FastAIAgent SDK classes and functions.

Core

  • Agent — Agent, AgentConfig, AgentResult
  • Chain — Chain, ChainResult, ChainState
  • LLMClient — LLMClient, LLMResponse, Message
  • Decisions API — LLMClient.decide() / adecide(), Predicate, Choice, Score, Option, Level, DecisionResult, PredicateAnswer, ChoiceAnswer, ScoreAnswer, Refusal; DecisionJudge; decision_tool

Tools

  • Tool — Tool, ToolResult, FunctionTool, RESTTool, MCPTool

Safety

  • Guardrail — Guardrail, GuardrailResult, no_pii, json_valid, toxicity_check
  • Guardrail primitives — run_guardrail(guardrail, data) (compute only, applies on_error), plane_guardrails_for_agent(agent_id), guardrail_from_policy_rule(rule). Use these when the runtime isn't fa.Agent.

Durability (v1.0)

  • interrupt / Resume — interrupt(reason, context), Resume(approved, metadata), InterruptSignal, AlreadyResumed
  • @idempotent — idempotent, IdempotencyError
  • Checkpointer Protocol — Checkpointer, SQLiteCheckpointer, PostgresCheckpointer, PendingInterrupt, Checkpoint
  • Resume entrypoints — Chain.aresume(execution_id, *, resume_value=Resume(...)), Agent.aresume(...), Swarm.aresume(...), Supervisor.aresume(...). All four runner types share the same atomic-claim contract.

See the full durability API reference for exact method signatures, the Checkpointer Protocol surface, and the agent_path hierarchy used by multi-agent topologies.

Observability

  • TraceStore — TraceStore, TraceData, SpanData, trace_context
  • Replay — Replay, ReplayStep, ReplayResult
  • OpenInference emitters — emit_guardrail(tracer, ...) / emit_evaluation(tracer, ...) open+stamp+close a child span on your tracer; set_guardrail_attributes(span, ...) / set_evaluation_attributes(span, ...) stamp a span you already own. evaluation.score is a 0..1 scale.

Intelligence

  • PromptRegistry — PromptRegistry, Prompt, Fragment
  • LocalKB — LocalKB, Document, SearchResult
  • PlatformKB — PlatformKB (hosted; thin client over /public/v1/knowledge-bases/{id}/search)

Evaluation

  • evaluate — evaluate, Dataset, Scorer, EvalResults

Platform

  • Connection — fa.connect(), fa.disconnect(), fa.is_connected. connect(export_traces=False) connects for policy/scorers/prompts without registering a span exporter or claiming OTel's global provider — for runtimes that already own one.

For detailed usage, see the individual module documentation pages linked above.