Email is where founders lose the most hours to work that follows rules. Triage, intro drafts, follow-ups, the reply that says the same thing the last forty replies said. This blueprint is a complete agent that takes that work: it reads the inbox, classifies what matters, drafts outreach, and answers threads on its own.
Everything you need to run it is on this page: the live console it ships with, its full anatomy, the system prompt, the source, and the setup. It deploys in an afternoon.
See it running
This is the console the agent ships with: an inbox where every AI action waits for your approval before it sends. Click into it, it is live.
The anatomy
An agent is not a prompt. It is a composition: a model, tools it is allowed to call, an integration that gives those tools reach, and guardrails that decide what it may never do alone. This is the whole machine on one card.
- Model
- claude-sonnet-4-6
- Tools
- send_intro_emailread_inboxauto_reply
- Integration
- AgentMail
- Trigger
- inbound email, or on request in chat
- Memory
- thread context plus contact history
- Output
- sent reply plus one short status line
- Guardrails
- no invented claims; asks one targeted question when input is missing
The system prompt
Twelve lines. Short on purpose: every rule the agent needs and nothing it could misread. Copy it as a starting point for any operations agent you build, the structure transfers.
The source, on disk
Three files carry the whole agent: the agent definition with its three tools, the environment it expects, and the setup. Walk them.
The daily loop
Once deployed, the agent runs the same four-beat loop on every message, all day. You only exist in the approve step of the console above.
Receive
Every inbound lands through AgentMail with full thread context.
Classify
Priority and intent before anything is written.
Act
Drafts, replies, and follow-ups within its guardrails.
Report
One status line per action, receipts in the console.
The business case
If you run an agency, this blueprint is also a product. Email operations is the first thing every client wants off their plate, and the numbers work at very small scale.
setup fee you can charge
retainer per client
email ops saved weekly
Ship it
The whole install is five commands and three keys. Run the list, and use the first-run ritual in setup.md to prove the loop end to end before you point it at a real inbox.
Deploy checklist
0 of 6That is the entire blueprint: one model, three tools, one integration, and guardrails that keep it honest. Built by hand it costs you an afternoon of cloning, keys, and deploys. Inside Ultron it costs you a sentence, because you can build this exact agent in the workforce and watch it come alive before you wire anything.
Build the Email Agent in Ultron
Start from this blueprint or your own words. Ultron reads the intent and picks the tools.
The model, the three tools, the prompt, the console. Assembled in front of you, nothing to wire.
Test it live against a real thread, then put it to work. You review a feed, it handles the inbox.