™An agent is only as good as the information it can see. Three things decide what it knows at any moment.
1. Context: what's on the desk
The context is everything the agent can see right now: its instructions, the request, the conversation so far, and the results of any tools it has used. Think of it as the papers spread across a desk.
The desk has a size limit. Modern models can hold a lot, but not everything, and very long contexts can make an agent slower and less focused. Good agents keep the desk tidy and only bring out what the current job needs.
2. Reference documents: the filing cabinet
Most business agents need to know things about your business: services, prices, policies, procedures, product details. Rather than putting all of it in the instructions, you give the agent a set of reference documents it can search.
When a question comes in, the agent looks up the relevant pages and brings just those onto the desk. Technical people call this retrieval.
The quality of these documents decides the quality of the answers. If your returns policy is out of date, the agent will confidently quote the old one.
3. Memory: notes it keeps between conversations
Some agents can save notes for later: a customer's preferences, a decision you made last week, a task still in progress. That's memory. Without it, each conversation starts from scratch.
Memory is useful, and it needs care. Decide what an agent should remember, check what it has saved from time to time, and be careful with personal information.
Why this matters more than the model
Most disappointing AI agents don't fail because the model is weak. They fail because the information is missing, outdated or messy. If your procedures only exist in someone's head, an agent can't follow them.
The good news: getting your key information written down and organized helps everyone. New hires, your team and your AI tools all benefit.