# Not every AI system needs planning, memory, tool autonomy, multi-agent coordi...
Canonical: https://social-archive.org/yena/4SYh8eLNOX
Original URL: https://x.com/itsharmanjot/status/2071660558244549097
Author: Harman
Platform: x
## Content
Not every AI system needs planning, memory, tool autonomy, multi-agent coordination, or a long reasoning loop. A helpful way to plan your AI system is to work backwards from the job: 1. *What is the outcome? Is it generating content, answering questions, classifying requests, or completing an action? 2. *Does it need current company knowledge? If yes, add retrieval or RAG. If no, a model with a good prompt may be enough. 3. *Does the output need to be reliable and machine-readable? Add structured outputs before adding more “intelligence.” 4. *Does it need to do something outside the chat? Add tool or function calling only when it needs to read from, write to, or trigger another system. 5. *Is the process predictable? Use a workflow. Fixed steps are usually easier to test, govern, and improve. 6. *Does the system genuinely need to decide the next step for itself? That is where an agentic loop starts to make sense. And the moment an AI system can access business data or take actions, production controls stop being optional: • scoped identity and permissions • observability and tracing • evaluations • fallback paths • approval rules for higher-risk actions The goal is not to build the most agentic system. It is to build the smallest system that can safely and reliably do the job. Start low in the stack. Add complexity only when the task earns it.