Caesar: you talk to one agent, and a whole team gets to work
Author: Adriano Orkisz
Almost every company that adopts AI ends up with several separate tools. One bot answers on WhatsApp, another writes posts and a third looks for clients. Each has its own panel and its own settings, and the owner is still the link between them all.
Caesar is our answer to that mess. We are building it at Imperador AI and we work with it ourselves. At the top there is a short video introducing the project.
How you use it
You open Caesar in the browser, hold down the button or the space bar and speak, for example: "find beauty salons in Fuengirola that don't have a website". Caesar answers by voice and hands the job to the right department. Esc interrupts it mid-sentence.
Only you talk to Caesar. The agents talk to each other on a shared channel, and you see who is doing what and how much it costs.
Departments, not loose bots
Caesar passes the work to the head of each department, who shares it out among their agents. Today we have:
- Sales: prepares the messages to leads on WhatsApp and keeps track of follow-ups.
- Leads: looks for businesses, checks their profiles and saves the results in the CRM.
- Marketing and social: a posting plan, texts and graphics for approval.
- Operations: work in the browser, email, calendar and reminders.
An agent is a row in the database: a prompt, a model, tools and a permission level. A new one is added from the panel, with no code deployment.
A person approves everything that goes outside
This is a rule you cannot get around with a setting. Sending a message, publishing, inviting or paying: every such action goes to the Approvals tab. You get a batch with one sentence per item, for example "3. Pizzeria Roma, no website, we propose a menu demo". You reply "yes 1 2 4, reject 3" and the rest moves on.
In addition:
- Limits: each task has a token budget and a maximum number of conversation rounds between agents. If it is exceeded, Caesar stops the work and asks you what to do. Departments also have their own daily budgets.
- Figures with a source: a separate control agent rejects a report in which a figure has no source in the collected data. A business's rating and its number of reviews come from a specific page, not from the model's memory.
- Payment block: the browser the agent drives is forbidden from paying in code. When it meets a CAPTCHA or a 2FA code, it asks you to take over.
Where you talk to it
In the browser and in an app for macOS and Windows. The same commands also work through Telegram and WhatsApp. Simple commands, such as "status", "cost today" or "yes 1 2 4", are handled by code without calling the model, so they cost nothing.
What it is built on
- Claude as the model in the conversation and in the agents' work, with prompt caching to bring the cost down.
- Voice: a Polish-language voice on our server, with the browser's speech recognition.
- Database and CRM in Supabase, with tasks in a queue served by a separate worker.
- A web app in Next.js and a desktop version in Tauri.
Where it stands
Caesar runs internally at our company and is not yet for sale. The idea is for it to become a product in two forms from the same code: a web app and a desktop app. Next up is lowering the daily cost, learning from sales results (which moves get a reply) and, later on, the phone: Caesar calling by itself in urgent matters, with a limit and quiet hours.
We are not giving a price or a date yet. If you want to hear first, or you are wondering which processes in your business such a team could take on, write to us. We will show you Caesar live.
Frequently asked questions
Does Caesar send messages to clients on its own?
No. Sending, publishing and paying always go through your approval.
Can I add my own agent?
Yes. An agent is a record in the database with a prompt, a model and tools, so you add it from the panel.
What happens if the agents get stuck in a loop or spend too much?
The task has a limit on tokens and rounds. When it is exceeded, the work stops and Caesar asks you for a decision.
