Skip to content

Guardrails, policies, personas & skills

These decorators capture the governance and identity layers of an agent.

Guardrails

Validate inbound requests and outbound responses. Each spec carries an action (block / warn / log) and optional severity.

from langstitch import input_guardrail, output_guardrail

@input_guardrail(description="Reject empty/oversized input.", action="block")
def non_empty(text: str) -> bool:
    return bool(text and 0 < len(text) <= 8000)

@output_guardrail(description="No secrets in responses.", action="warn")
def no_secrets(text: str) -> bool:
    return "BEGIN PRIVATE KEY" not in text

Run get_input_guardrails() before the LLM and get_output_guardrails() after, honoring each spec's action and severity.

Business policies

Organizational rules evaluated by priority (highest first). Short-circuit on the first non-allow decision.

from langstitch import business_policy

@business_policy(priority=100, description="Deny refunds over policy limit.")
def refund_limit(context: dict) -> dict:
    amount = context.get("amount", 0)
    return {"decision": "deny" if amount > 1000 else "allow"}

from langstitch import get_policies

for policy in get_policies():          # already sorted by priority
    decision = policy.fn(context)
    if decision["decision"] != "allow":
        break

Personas

A persona is an agent identity / system prompt, selected by name at context-build time.

from langstitch import persona

@persona(role="assistant", tone="helpful, concise")
def assistant() -> str:
    return "You are a helpful LangStitch support assistant."

Skills

A skill is a reusable capability. Like every decorator it works bare or parameterized:

from langstitch import skill

@skill
def echo(text: str) -> str:
    return text

@skill(name="search", tools=["web"], tags=["retrieval"])
def web_search(query: str) -> list[str]:
    ...

Tools

A @tool is a callable an LLM can invoke. Gate access with roles and select by tags/name/role at context-build time.

from langstitch import tool

@tool(tags=["billing"], roles=["agent"], input_schema={"invoice_id": "str"})
def lookup_invoice(invoice_id: str) -> dict:
    """Fetch an invoice by id."""
    ...

Designers in LangTailor

The LangTailor IDE ships visual designers for skills, guardrails, business rules, and personas that export to exactly these decorated modules.