Multiple AI email agents, routed by topic

What an agent is

An agent is the set of instructions behind a reply. It answers four questions for the AI: who are you writing as, what should each reply achieve, what voice should it be in, and what are you allowed to point people to. In inboxr an agent is made of:

• Role and goals. A plain-language description of the job, for example "You handle billing questions for a small design studio. Confirm the invoice number, answer from our billing policy, and never promise a refund."
• Tone. The voice it writes in, from formal to friendly.
• Links. Call-to-action links it may include, each with a note on when to use it.
• Knowledge. Facts attached to this agent only, on top of your shared knowledge base.

One mailbox can have several agents. That is what lets a single shared address, such as a hello@ or support@ inbox, answer a pricing question and a password question in two different ways.

How routing works

1. The email is classified. Before any draft is written, inboxr reads the message and assigns it one topic: billing, sales, support, account, or general. Anything that does not fit is marked as other.
2. The matching agent is chosen. If you have assigned an agent to that topic on the mailbox, it writes the draft.
3. Otherwise the default agent writes it. Every mailbox has one default agent, which handles anything not routed elsewhere. You choose which agent is the default.

The topic is shown as a badge on each item in the review queue and can be used as a filter, so you can work through all billing drafts in one go.

Sample data. Names and messages are fictional.

Why separate agents help

Different goals

A sales reply should move toward a next step. A billing reply should be exact and careful. A support reply should resolve the problem in one message. Writing all three into one set of instructions produces replies that are vague about all of them.

Different links

Your sales agent can carry a booking link. Your support agent can carry help articles. Neither will offer the other's links, because links belong to the agent.

Different knowledge

Knowledge can be attached to one agent so only that agent uses it. You can also limit an agent to a chosen set of knowledge articles. With nothing selected, an agent uses all of your active knowledge, which is the default.

Learning that stays relevant

When inboxr picks past replies to learn from, it prefers examples written by the same agent and on the same topic. Corrections you make to billing replies improve billing replies first. See learning from your edits.

Setting it up

1. Create one agent with a clear role and tone and make it the default for the mailbox.
2. Run in co-pilot for a while and watch the topic badges in the queue.
3. When one type of email keeps needing the same edits, create an agent for it and assign it to that topic.
4. Give the new agent its own links and, if it needs them, its own knowledge.

Keep instructions short and specific. "Be warm, confirm the order number, and offer the returns link when relevant" works better than a long paragraph of general advice.

What stays the same across agents

Agents change how a reply is written, not what happens to it afterwards. Every draft, from any agent, follows the mailbox's mode: held in the review queue in co-pilot, or scored and checked before sending in autopilot. The mailbox's signature, reply-from name, send caps, and ignore rules apply to all of them.

Who it is for

Teams with one front-door address that receives several kinds of email, and anyone who has noticed that their sales voice and their support voice are not the same. If your mailbox only ever gets one kind of question, one default agent is all you need. For worked setups, see the billing questions and sales inquiries use cases.