Introspection

describe() dumps the active rule set and the fact schemas as plain data:

TrafficAdvice.describe()
# {
#   "facts": [{"name": "Light", "fields": {...}}],
#   "rules": [
#     {"name": "green",   "predicate": {"type": "Eq", "field": "color", "value": "green"},
#      "priority": 3, "is_default": False},
#     {"name": "yellow_safe", ...},
#     {"name": "stop",    ...},
#     {"name": "fallback","predicate": null, "priority": 0, "is_default": True},
#   ],
# }

This is the canonical way to:

Render a rule set in an external tool or UI
Diff two versions of a rule set
Persist the rule set to a store
Feed it back into an LLM so the model knows exactly which cases are already handled

Combined with Predicate.from_dict, predicates round-trip cleanly, so you can load a previously stored rule set and re-evaluate it.

Feeding the rule set to an LLM

When the engine's @Default calls an LLM, passing type(self).describe() in the system prompt tells the model precisely which cases are already covered by rules - preventing it from reasoning about inputs that will never reach it:

import json

@Default
async def llm_fallback(self, ticket: Ticket) -> Team:
    rules_schema = json.dumps(type(self).describe(), indent=2)
    agent = Agent(
        "anthropic:claude-haiku-4-5",
        result_type=Team,
        system_prompt=(
            "You are a support ticket classifier. "
            "These deterministic rules already handle common cases - "
            "you only receive tickets that matched none of them.\n\n"
            f"Existing rules:\n{rules_schema}"
        ),
    )
    result = await agent.run(f"Subject: {ticket.subject}\n\n{ticket.body}")
    return result.output

Without this, the LLM and rules engine are two separate systems with no shared understanding. With it, they form one coherent pipeline.

Feeding the rule set to an LLM​

Feeding the rule set to an LLM