^ This. Also, Ive had amazing success focusing on deterministic guardrails! My legal bot journey taught me that safety is everything. I compared two methods:

• Microsoft Semantic Kernel SDK provides fantastic reliability thru its structured planners.
• Graph-based logic offers better cost control vs open loops. Honestly, I love the energy here! Just watch those token costs on OpenAI GPT-4o API if it starts looping.

Honestly, agents are way overrated right now. I had issues with loops because everyone trusts the reasoning too much. Unfortunately, even OpenAI GPT-4o API breaks if you dont force strict outputs. I tried LangChain Framework Python but it was not as good as expected for logic. Self-reflection is just another chance for it to fail tbh... Use deterministic state machines or you will be babysitting it forever.

> Is it more about self-reflection or is there a better way to handle planning... Jumping in here because I love the energy! Honestly, its all about hierarchical planning. Instead of letting one agent spin its wheels, use a manager-worker setup. I found that CrewAI Framework is amazing for this because it lets you bake in a process that forces a handoff. Tbh, self-reflection is okay but it gets expensive fast and sometimes just loops the same mistake over and over. For the actual brain, Anthropic Claude 3.5 Sonnet API has been fantastic for my legal workflows lately... its super sharp on logic and handles long contexts without getting as lost as others. If you're worried about costs, try routing the simpler validation tasks to Groq Llama 3 70B since its insanely fast and cheap. Definitely helps keep the agent from looping forever because a manager agent can just kill the task if it sees the same output twice!