Framework Safety Enforcement

approval and guardrails YAML fields fail closed on frameworks that cannot enforce them

PraisonAI refuses to start a run that would silently drop a safety gate.

​

Quick Start

1. Gate on the native framework

   framework: praisonai honours approval: and guardrails: natively.

   agents.yaml

   framework: praisonai
   topic: Compliance review
   roles:
     compliance_agent:
       role: compliance-checker
       goal: Review the contract for policy issues
       backstory: You are a careful compliance reviewer.
       approval:
         backend: console
       guardrails: strict
       tasks:
         review:
           description: Review the contract and list policy issues.
           expected_output: A bullet list of issues.
   

   praisonai agents.yaml
   
2. Same YAML on CrewAI refuses to start

   Change only framework: crewai. The CrewAI and AutoGen adapters cannot pass these fields through, so the run stops before any tool executes:

   ValueError: framework='crewai' cannot enforce safety field(s) ['approval', 'guardrails']
   for agent 'compliance-checker'. Use framework: praisonai, or remove the field(s) to run ungated.
   
3. Pick a fix

   Either switch back to framework: praisonai to keep the gate, or delete the approval: / guardrails: keys (or set approval: false) to run ungated on purpose.

Behaviour change. Previously these fields were dropped with only a log WARNING, and tools ran ungated. Any YAML that combines approval: or guardrails: with framework: crewai or framework: autogen now fails at startup.

---

​

When does the error fire?

YAML value	Result on crewai / autogen
approval: false or approval: null	Warns, runs (gate disabled)
approval: {enabled: true, backend: auto, approve_all_tools: false} (the --trust shape)	Warns, runs (auto gates nothing)
approval: {enabled: false, ...}	Warns, runs (gate disabled)
approval: {backend: none}	Warns, runs (none gates nothing)
approval: true	ValueError (defaults to the console backend)
approval: {backend: console} (or slack, webhook, …)	ValueError
approval: {typo: 1} (unparseable)	ValueError — invalid config fails closed
guardrails: null or guardrails: ""	Warns, runs (no gate)
guardrails: strict or any non-empty value	ValueError

Non-safety fields an adapter does not read (for example memory on CrewAI) still only log a warning.

​

Who is affected

framework:	Honours approval / guardrails	On an active gate
praisonai	✅ Yes	Gate enforced
crewai	❌ No	ValueError
autogen, autogen_v2, autogen_v4, ag2	❌ No (only llm is supported)	ValueError

YAML files that never set approval: or guardrails: are unaffected.

---

​

How to migrate

Original YAML	Keep the gate	Remove the gate	Disable the gate
framework: crewai + approval: {backend: console}	Change to framework: praisonai	Delete the approval: key	Set approval: false
framework: autogen + guardrails: strict	Change to framework: praisonai	Delete the guardrails: key	Set guardrails: null

---

​

Best Practices

Use framework: praisonai for gated tools

Approval and guardrails are only enforced on the native framework. If a role touches shell, file, or network tools, run it on praisonai.

Disable explicitly, not by accident

When you really want CrewAI or AutoGen to run ungated, write approval: false or remove the key so the intent is visible in review.

Treat the error as a config check

The error fires at startup, before any agent runs. Add praisonai agents.yaml to CI to catch dropped gates early.

---

​

Related

Approval

Human-in-the-loop approval for tools

Guardrails

Validate agent output