A team of AI agents running a real business, 24/7.
Before we sell agents, we run our own operations on them. Sales agents answer for three businesses, a voice agent makes calls on request, and a group of internal agents keeps the stack healthy with shared memory, scheduled jobs and monitoring. People approve anything that matters.
The problem.
- Messages arrive on many channels, at all hours, for more than one business.
- Operational chores (renewals, deploys, monitoring, follow-ups) pile up and fall through the cracks.
- A single chatbot can't do this safely. It needs separate identities, shared context and clear limits on what each agent may do.
What we built.
Sales agents
Three separate agents with their own identity and knowledge: Regina (our demo agent), Valeria for our real-estate portal Casa y Predio, and Fernanda for the restaurant. All three answer website chat; Regina and Fernanda also answer Instagram and Facebook.
Lead cards
When a conversation shows real interest, the agent asks once for contact details and sends one lead card (need, channel, next step) to a person by chat and email. No alert per message.
Outbound voice and WhatsApp
A voice agent places calls (and WhatsApp messages where enabled) only on explicit request, within calling hours, and reports a summary of each call.
Internal ops agents
A team of agent profiles with roles (operations, coding, QA, social, research), each with its own memory and access to a shared knowledge wiki.
Scheduled jobs
Recurring jobs handle reconciliation, health checks and reports. Each one is listed in an inventory with an owner and schedule, and stale ones get removed.
Monitoring
Watchdogs check public endpoints, models and services. They alert only after consecutive failures, and an agent fixes the problem before a person is pinged.
How it fits together.
Ops agents read and update the same wiki the sales agents are grounded on, so a fix made once applies everywhere.
Simplified diagram. Hostnames, keys and internal addresses are left out on purpose.
The stack.
- Agent runtime with per-agent profiles, memory and tool permissions
- Shared knowledge wiki as the single source of operating facts
- LLM inference via hosted endpoints and our private cluster
- Meta messaging APIs (Instagram, Facebook) and web chat with signed sessions
- Voice agent platform with telephony, plus WhatsApp Business
- Scheduled jobs and watchdogs with alert thresholds
How we test and roll back.
- Each agent change ships with tests. Price guards are tested with real numbers and with tricky non-prices such as list numbers, so correct replies are not blocked and invented prices still are.
- Probe conversations use a recognizable test prefix and are deleted afterwards, so they never count as leads.
- Production changes to client agents follow one routine: dated backup, syntax check, restart, health check, a real conversation, and a note in the wiki.
- Rollback means restoring the dated backup and restarting the affected service. Our backups are kept until the change is proven, then pruned.
Where people stay in control.
- People close sales, quote anything unusual, approve spending and sign agreements. Agents point to a human WhatsApp and email for that.
- Agents never make up prices: replies with an unpublished amount are blocked and replaced with a safe answer.
- Outbound calls and messages go out only on explicit instruction and within set hours.
- Ad accounts, payments and anything public under a person's name need that person's approval.
What the client gets.
- Agents with your brand's names, each limited to its own business and knowledge
- Lead cards to the person who sells, not a flood of alerts
- Monitoring and scheduled jobs for the systems the agents touch
- A monthly review of conversations, guards and costs
Comparable packages
$1,500 USD/mo
One-time setup $6,000
Sales agent on calls, texts, web and Facebook/Instagram.
$750–3,000 USD/mo
After an Automation Build
Monitoring, scheduled jobs and ongoing improvements.
Prices from our public US price list. Telephony, messaging, AI usage and payment fees are passed through at cost. Final scope in writing.
Status and what we don't publish.
Verified November 2026: these agents and jobs run every day across the businesses we operate. We don't publish conversation or lead counts here, because our own numbers include internal testing and would not be a fair benchmark for your business.
Want this for your business?
Book a 20-minute call. We look at your channels, your systems and where an agent pays off, and you leave with a plan.