add-guardrails

# Add Guardrails to AG2 Agents You are an expert at adding safety guardrails to AG2 multi-agent systems. When the user wants to add safety constraints: ## 1. Understand the Requirements Ask the user: - What should be blocked? (PII, off-topic responses, harmful content, specific patterns) - What should happen when a guardrail triggers? (redirect to another agent, terminate, sanitize) - Is pattern matching sufficient (RegexGuardrail) or does it need LLM judgment (LLMGuardrail)? ## 2. LLM-Based Guardrail For complex, nuanced safety checks: ```python from autogen import ConversableAgent, LLMConfig from autogen.agentchat.group import AgentTarget, TerminateTarget from autogen.agentchat.group.guardrails import LLMGuardrail llm_config = LLMConfig( {"model": "gpt-4o-mini", "api_key": os.environ["OPENAI_API_KEY"]} ) # Create a moderator to handle violations moderator = ConversableAgent( name="moderator", system_message="A guardrail was triggered. Politely explain that the response was filtered and ask the user to rephrase.", llm_config=llm_config, human_input_mode="NEVER", ) # LLM evaluates whether the response violates the condition guardrail = LLMGuardrail( name="relevance_check", condition="The response is off-topic, contains speculation not based on data, or makes claims without evidence", target=AgentTarget(moderator), llm_config=llm_config, # Required — the guardrail needs its own LLM ) # Register on the agent whose output should be checked assistant.register_output_guardrail(guardrail) ``` ## 3. Regex-Based Guardrail For fast, pattern-based checks (no LLM cost): ```python from autogen.agentchat.group.guardrails import RegexGuardrail # Block SSN patterns pii_guardrail = RegexGuardrail( name="ssn_filter", condition=r"\b\d{3}-\d{2}-\d{4}\b", target=AgentTarget(sanitizer), ) assistant.register_output_guardrail(pii_guardrail) # Block email addresses email_guardrail = RegexGuardrail( name="email_filter", condition=r"\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b", target=AgentTarget(sanitizer), ) assistant.register_output_guardrail(email_guardrail) # Block credit card numbers cc_guardrail = RegexGuardrail( name="cc_filter", condition=r"\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b", target=TerminateTarget(), ) assistant.register_output_guardrail(cc_guardrail) ``` ## 4. Multiple Guardrails Stack multiple guardrails on an agent — they are checked in registration order: ```python # Check PII first (fast, no LLM cost), then content quality (LLM-based) assistant.register_output_guardrail(pii_guardrail) # Regex — fast assistant.register_output_guardrail(email_guardrail) # Regex — fast assistant.register_output_guardrail(relevance_guard) # LLM — slower but thorough ``` ## 5. Guardrails with Group Chat ```python from autogen.agentchat import run_group_chat from autogen.agentchat.group.patterns import AutoPattern # Register guardrails before creating the pattern researcher.register_output_guardrail(pii_guardrail) writer.register_output_guardrail(relevance_guard) result = run_group_chat( pattern=AutoPattern( initial_agent=researcher, agents=[researcher, writer, moderator], user_agent=user, group_manager_args={"llm_config": llm_config}, ), messages="Research and write a report on...", max_rounds=15, ) ``` ## 6. Rules - `LLMGuardrail` requires its own `llm_config` parameter — it won't work without one - `RegexGuardrail` uses Python regex syntax — test patterns before deploying - Register regex guardrails before LLM guardrails for efficiency (fail fast on cheap checks) - The `target` determines where execution goes when the guardrail triggers - Use `TerminateTarget()` to stop the conversation on critical violations - Use `AgentTarget(moderator)` to redirect to a handler agent - Guardrails check the agent's OUTPUT, not input — they fire after the agent responds