We have deployed hubs for jewellery retail, manufacturing, and IT operations. Different industries, very different teams. The same three agent shapes show up every time.
Type 1 is a Tool. The user presses something and an outcome appears. A festival creative gets generated. A vendor email gets drafted. A close-cycle export lands as a file. The interface is closer to a button than a chat. Tools live in Claude Code, in-app buttons, or simple internal scripts.
Type 2 is Q&A. The user asks a question in plain language and the hub answers using your data and policies. Conversational, but bounded by what the hub knows. Q&A agents live on Telegram or Slack because that is where operators already are. They read fluently in English, Hindi, French, German, or whatever language the team prefers.
Type 3 is Background. No one asks. The agent runs on a schedule. It reconciles overnight. It scans for anomalies. It drafts a Monday board pack. It pushes a daily research digest. The output lands in email or a Telegram channel before anyone logs in.
Mixing the three is where projects fail. We see teams build a Q&A bot and call it a tool. We see teams automate a Tool when a Q&A would have been more flexible. We see teams write Background agents that need a human to invoke them, defeating the point.
The right composition for one team usually has all three. Marketing needs a Strategy Q&A on Telegram, a Studio Tool in Claude Code, and a Research Background that lands a digest each morning. Finance needs a Reconcile Q&A, a Closer Tool, and an Anomaly Background. The shape repeats; the surfaces stay close to where the team already works.
Knowing which type you are building changes everything downstream. Tools need explicit input contracts. Q&A needs broad knowledge plus narrow guardrails. Background needs schedules, deduplication, and escalation paths.
When we scope an engagement, the first conversation is which agents go where in this matrix. Until that picks up, nothing else does.