AI systems

Multi-agent AI systems for real business work

Specialist agents with written skills, checked output, and live connections to the tools a business runs on. I design, build, and run these systems, so the calls get answered, the leads get logged, and the meetings get booked without anyone chasing them.

Book a call
AJulio Cornavaca
WorkflowsAgent mapAI costsBuilderSearch the agent map...
Active workflows by space
48
Last 30 days
48workflows
ContentSalesOperationsClientsFinanceGeneral
Runs by space
6,214
Last 30 days
Content1,980
Sales1,420
Operations1,260
Clients640
Finance520
General394
Tasks by agent
6,214
Arthur1,920 tasks
Maya1,340 tasks
Emily1,105 tasks
Rita870 tasks
8 more979 tasks
Triggers
6
ScheduleWebhookFormCallEmailChat
Failing runs
2need attention
Skills updated
3this week
Live connections
7tools
Triggers
Website formWebhook
Inbound callPhone line
Daily 7:00 runSchedule
New bookingCalendar event
EmailInbox
Agents
A
ArthurChief of staff
E
EmilyContent
M
MayaVoice and booking
R
RitaOperations
+ 8 more agents
Skills
lead-intakeSkill
call-scriptSkill
booking-rulesSkill
brand-voiceSkill
content-qaSkill
weekly-reportSkill
Tools
HubSpotCRM
RetellVoice
Google CalendarScheduling
WebflowWebsite
OpenRouterModels
MakeScenarios
CloudflareWebhooks
Maya · 2 triggers · 3 skills · 4 toolsActive path Other connections
Estimated spend$384
Tokens38.4M
Runs6,214
Cache hit rate21%
Via OpenRouter ▾
Spend vs monthly limit
$384 of $600
This month
Remaining$216
Spend by model
$384
claude-opus-5claude-sonnet-5gpt-5gemini-2.5-flashllama-4-maverick
Models

Runs, tokens, cache hit rate, and cost per model, this month.

ModelRunsTokensCacheCost
claude-opus-5Anthropic6409.8M18%$168
claude-sonnet-5Anthropic2,11014.2M24%$121
gpt-5OpenAI8206.1M12%$64
gemini-2.5-flashGoogle1,9305.4M31%$22
llama-4-maverickOpen-source7142.9M0%$9
← Lead intakecornavaca.com/api/lead-intakeRun ▸
README.md

Takes a new website lead, creates the contact in HubSpot, hands the call to Maya, and books a slot in Google Calendar.

Every result is reviewed by Arthur before it is sent.

Output
200 OK
1{
2"status": "created",
3"contact": {
4"name": "Dana K.",
5"source": "website form",
6"owner": "Maya",
7"lifecycle": "lead"
8 },
9"next_step": "call booked Thu 10:00",
10"checked_by": "Arthur"
11}
Dashboard36 runs today
Lead intake22 runs today
Booking18 runs today
Content queue27 runs today
Weekly report9 runs today
Call summaries14 runs today
Client portal11 runs today
Invoices6 runs today
Reviews8 runs today
Follow-ups19 runs today
12
Specialist agents
7
Live connections
4
Parts in one system
How it works

How the system works

Four parts turn separate AI agents into one working system. Work comes in on a schedule, from an event, or from a direct request. Orchestration plans it and checks it, skills tell every agent how the work should be done, and model routing sends each task to the model that fits it. The result is work that gets done the same way every time, with a record of every step.

Most AI setups stop at a chatbot: you ask, it answers, and nothing happens unless you're there to push it. A system works differently. Work starts on its own, every task follows the same written rules, and one agent is accountable for checking the result before anything goes out.

The diagram shows how the pieces connect. Automation brings the work in. Orchestration decides who handles it and reviews it. Skills keep every agent consistent, and model routing matches each task to the right model, so hard problems get the strongest models while routine work stays fast and inexpensive.

Cron runsWebhooksYou ask SCHEDULEDEVENTSDIRECT ASK AutomationSkills Skill library Model routing Orchestration FrontierFastOpen-source Complex reasoningRoutine volumePrivate work
The four parts

What each part does

01

Orchestration

One agent plans the work, checks the result, and logs every step.

02

Skills

Rules written once and followed by every agent.

03

Automation

Work starts on a schedule or the moment something happens.

04

Model routing

Each task runs on the model that fits it, with cost tracked.

Part 01

Orchestration

What it is

Orchestration is the layer that turns separate AI agents into one working system. Without it, you're the coordinator: picking which tool to use, copying results from one into the next, and checking everything yourself. With it, one agent does that job. In my setup that's Arthur, the chief-of-staff agent. He receives every request, breaks it into steps, decides what can run at the same time and what has to wait, and hands each step to the specialist agent built for it. Think of him as the project manager: the specialists do the work, and Arthur makes sure the right work happens in the right order.

Why it matters

A single AI doing everything guesses at work it isn't set up for. With a coordinator, each part of the job is handled by an agent built for it, and one agent is accountable for the final result.

How it works

Every request passes through one coordinating agent. It plans the work, reviews the output against the original request, and releases only approved work. Every step is logged, so there's a full record of what happened.

1Receive2Plan3Delegate4Pass context5Recover6Check7Record
A request comes in.
Request inNew website lead
Result deliveredChecked and sent
OrchestratorWaiting for work
1Look up the companyparallel
2Create the CRM recordparallel
3Write the replyneeds 1
Answers the request
Facts match the sources
Follows the rules
Research agentWeb search
Idle
Action agentCRM
Idle
Writing agentEmail draft
Idle
Run log
Part 02

Skills

What it is

Written instructions the agents load before they work: brand voice, rules, guardrails, and step-by-step workflows. Think of it as an operations manual the agents actually read every time.

Why it matters

Without skills, you correct the same mistake again and again. With them, you fix it once, and every agent follows the new rule from then on.

How it works

Each skill is a file covering one kind of work, such as brand voice or quality checks. Agents load the skills that apply before they start, and an update to a skill takes effect everywhere it's used.

Agents using skills
12
Skill library
brand-voicecontent-qalead-intake
Tasks run with skills
E
Emily · Content
Passed
Skills appliedbrand-voicecontent-qa
Blog draft · reviewed by Arthur2.4s
Client email · reviewed by Arthur1.8s
Carousel copy · reviewed by Arthur3.1s
◆
brand-voice · skill
Toneplain, direct, first person
SpellingUS
Claimsa number or a source
Neversales framing, filler
Used byEmily, Maya, Rita
--- name: brand-voice description: How every agent writes --- # Rules - Use US spelling - Every claim gets a number or a source - No sales framing, no filler

Loaded by every agent

Agents read the skills that apply before they start a task. No retraining, no reminders.

Visible on every run

Each run shows which skills were applied, so you can see why the output looks the way it does.

Rules in plain language

Brand voice, guardrails, and workflows, written the way you'd explain them to a new hire.

Written once

Skills are files. Fix a rule once and the change reaches every agent that uses it.

Part 03

Automation

What it is

The part that starts work without anyone asking. Some work runs on a schedule, and some starts the moment something happens.

Why it matters

Agents that wait for a prompt are just a faster chatbot. With automation, follow-ups, reports, and calendar updates happen while you're doing something else.

How it works

Scheduled jobs run recurring work at set times. Event triggers, called webhooks, start work the moment something happens, like a form submission, an inbound call, or a new booking.

Scheduled work runs on its own.
SCHEDULEDWeekdays · 07:00Waiting
CloudflareCron trigger
MakeScenario runs
HubSpotPull new deals
GmailSend daily report
EVENTWebhook · form submittedWaiting
WebflowForm submitted
MakeScenario runs
RetellCall the lead
Google CalendarBook the meeting
Every run logged · trigger, steps, time
Part 04

Model routing

What it is

Choosing which AI model handles each task. Not every task needs the most powerful, and most expensive, model.

Why it matters

Using a top-tier model for everything is slow and costly. Using a cheap one for everything hurts quality. Routing gets both right.

How it works

Agent calls run through OpenRouter, which connects to many models from one place. Complex reasoning goes to frontier models, routine high-volume work goes to fast, low-cost models, and private work goes to self-hosted open-source models. Token usage is metered on every call.

Each task goes to the model that fits it.
Tasks
Models
ComplexDraft the Q4 strategy memo
High volumeTag 400 support emails
Private dataSummarize a client contract
OpenRouterWaiting for a task
Claude Frontierclaude-opus-5 · 0 tok
Gemini Fastgemini-2.5-flash · 0 tok
Llama Self-hostedllama-4-maverick · 0 tok
Metered on every call0 calls · 0 tokens
ClaudeGeminiLlama
For companies, business owners & executives

Business owner? Leading a team and need AI? I’m here to help.

AI can take work off your team’s plate. The hard part is knowing what to hand off, which tools to use, and how to make them work together. I help you figure that out and build it, and I’ll show you what AI can take over.

You don’t know where to start.

Every tool promises everything, and you can’t tell what’s worth paying for. We look at where your team’s hours go and pick the one system that pays off first.

Don’t know how to connect your tools or automate the work?

ChatGPT here, an automation there, and your team stitching it together by hand. One coordinating agent connects your tools, so the work moves on its own and every step is checked.

You’re scaling and need AI fast?

Volume is climbing and hiring can’t keep up. We put agents on the work that’s piling up first, with written rules and a record of every task so quality holds as you grow.

Find what to hand offOr send me your project details
Julio Cornavaca
Free guide · AI 101

AI agents, explained in 60 seconds.

What an AI agent is, how it reaches your apps, what it costs, and which kind fits you.

AI 101 free guide cover: AI agents explained in 60 seconds Get the free guide