The story
People ask this a lot, usually after an agent does something that feels clever. Is the cleverness in the agent, in the AI company, in the instructions? The honest answer is: in four places, and most of it is not where people point.
The one plain idea
An agent's results come from four stacked sources. (1) The model: patterns learned from huge amounts of text during training, stored as numbers called weights. (2) The instructions: files and prompts that tell it who it is and how to work. (3) The tools: the things it can touch. (4) The feedback: results coming back each loop, which let it fix mistakes. Take away any one and the agent gets much worse.
The kitchen-table picture
A new hire. The model is their education and raw ability. The instructions are the onboarding handbook. The tools are the laptop and the system logins. The feedback is the manager saying 'not quite' and the customer's reaction. A brilliant hire with no handbook, no laptop and no feedback does very little.
Worked example
Same model, three setups. Setup A: no tools, no instructions. Ask 'what is the status of order 4471?' It guesses or says it cannot know. Setup B: add an order-lookup tool. It now finds the real status. Setup C: add an instruction file saying 'always give the delivery date first, and be brief'. Same answer, now in the shape the business wants. The model's weights never changed across A, B, C. Only the instructions and tools changed, and the behavior changed a lot.
No-code exercise
Pick a task from your own job, such as answering a customer email. List what an AI would need under each of the four headings: what knowledge it must already have, what instructions you would hand it, what tools it must touch (inbox, order system), and how it would learn it got something wrong. Circle the weakest of the four. That is where an agent would fail first.
Self-check
- Name the four sources of an agent's intelligence. Answer: The model (weights), the instructions, the tools, and the feedback loop.
- In the order-status example, did the model get smarter between setups A and B? Answer: No. It got a tool. The weights stayed the same.
- Why does that matter to a business owner? Answer: Most of what you can improve yourself is in instructions, tools and feedback. You cannot retrain the model, but you can write a better handbook.
Sources
- Anthropic: the foundation of agentic systems is a model augmented with retrieval, tools and memory, and agents rely on environment feedback ('ground truth') at each step. https://www.anthropic.com/engineering/building-effective-agents General knowledge / our framing: How training stores knowledge in weights, and the four-part breakdown, are general knowledge and our own teaching framing, not a quote from any one source. Different companies would slice it differently.