What is context in an AI agent, and where does it come from?
Context is everything you feed a model so it can actually do a task — not the instruction, the information: documents, live web data, structured records, the tools it can call, and the systems where it stores and retrieves. As the models themselves commoditize, the quality of that context is the part that compounds. Pick almost any agent that fails in production and the cause isn't the model; it's one of these five sources being missing, stale, or badly represented.
Documents
The unstructured pile: PDFs, slide decks, contracts, reports. Most enterprise knowledge is locked in these, and most of it is read badly by older tools. Getting a document into a form an agent can actually use, accurately, is its own hard problem.
The web
Live information the model was not trained on. The ability to search and crawl so the agent is working from what’s true today, not what was true at training time.
Structured data
The clean stuff already sitting in SQL databases and warehouses: numbers, records, transactions. The agent needs a path to query it in natural language instead of waiting on a human to write the report.
Tools the agent can call
The software it can reach through connectors like MCP — Confluence, Salesforce, your internal systems. This is where context stops being a lookup and becomes an action: the agent doesn’t just read, it operates the tools.
Systems of record
Where the agent stores and updates state: a CRM entry, a Notion page, an internal wiki. Agents need somewhere durable to act on and retrieve data, not just read it once.
Why it matters
The framing to remember: stop asking which model is best and start asking what context you’re giving it. The model is converging across providers. The context is yours, and it’s the part that compounds.
From the conversation
This explainer is drawn from these episodes — each carries its full transcript.
