Servicesservicemulti-agent automation

AI agents that work as a team.

One model alone guesses. A team of agents plans, looks things up, acts on your systems and checks the result before it reaches a customer. Each step goes to the model that does it best, for the lowest cost. Designed, built and monitored by the same person.

  • A team of agents designed around your process
  • Connected to your systems: ERP, WhatsApp, email, spreadsheets
  • A panel to follow every run, its cost and its result
  • You approve what matters; the rest runs on its own
task: “Answer 240 support tickets”live
Planner
workingClaude

Split into 3 queues: refunds, tracking, other

♪♫
Research
waitingGPT · RAG

—

Operator
waitingLlama · local

—

Reviewer
waitingGemini

—

orchestrator
run.logstep 1/8
0.4splanner › Split into 3 queues: refunds, tracking, other
tokens
1,840
cost
US$0.06
models
1
scoreeach note is one model call
planner
research
operator
reviewer

A simulated run, step by step. Each note on the score is one model call.

01 · Where it fits

Work that follows rules but still needs judgment.

If someone on your team spends the day reading, looking up and copying between systems, there's probably an agent team for it.

01

Customer support

Reads the ticket, checks the order in the ERP, drafts the reply in your tone and only asks a person when it's delicate.

researchoperatorreviewer
02

Documents into data

Invoices, quotes, contracts and PDFs, scans included: extracted, matched against your records and registered in the system.

researchrevieweroperator
03

Reports that explain

Pulls numbers from every channel, compares periods and writes what changed and why, in plain language, on a schedule.

operatorresearchplanner
04

Sales on WhatsApp

Answers leads at any hour, qualifies them, checks stock and price, and hands the hot ones to your team with the context.

planneroperatorreviewer
05

Back-office routines

Marketplace questions, stock syncs, order follow-ups, overdue payments: the repetitive work that eats the team's day.

planneroperator
06

Company knowledge

An assistant that knows your policies, manuals and history, answers the team with sources and never makes up a rule.

researchreviewer

02 · Anatomy

Five pieces, each one under your control.

Behind the animation there's plain configuration. You can read it, change it and audit it.

orchestrator.yaml
1orchestrator:
2 model: claude # plans and delegates
3 goal: "answer pending tickets"
4 max_steps: 12 # never loops forever
5 on_doubt: ask_human

03 · Routing

The right model for each step.

The expensive model only thinks where thinking pays off. Simple steps go to cheaper or local models, and nothing is tied to a single provider.

Plan and decide
the hardest reasoning
Claude · GPT
100
Read documents and images
vision and long text
GPT · Gemini
46
Review and double-check
fast second opinion
Gemini Flash
18
Search your knowledge
cents per thousand documents
embeddings
6
Look up and update systems
runs on your server, data stays in
Llama · local
≈0

Illustrative proportions (most expensive step = 100). Real prices vary by provider and volume.

04 · Under control

Autonomous where it's safe. Yours where it counts.

Every agent team comes with a control room: what it did, what it's waiting for, how much it spent and how well it's doing.

control.room·run #1042running

A person approves what matters

Refunds, discounts, sensitive replies: the run pauses until you say yes.

15% discount on a quote#318

Above the 10% limit

ApproveEdit
Reply to a complaint#2207

Upset tone flagged by the reviewer

ApproveEdit
queue clear

Every step is logged

Who did what, when and for how much. Nothing happens in a black box.

  1. plannerRead ticket #4821: refund request
  2. researchFound the policy: 7 days to return
  3. operatorOrder delivered 3 days ago, first refund
  4. reviewerRefund R$ 189,90 needs approval
  5. youApproved by you
  6. operatorRefund issued through the payment API
  7. operatorCustomer notified on WhatsApp
  8. reviewerRun closed and logged

Budget with a ceiling

Limits per run and per day. If something goes off the rails, it stops before it costs you.

US$12.40/ 20 per day
ceiling
7 days agothis run: US$ 0.004

Tested on real cases

Each change runs against your real cases before going live. Each failure becomes a new test.

0/120real cases · v1.4
    policy100%
    tone98%
    data100%
    handoff96%
    0 failed, fixed before release testing

    Sample run with fictitious data.

    05 · How we build it

    From the first case to running on its own.

    Autonomy is earned step by step. The agents take on more of the work only after proving they can, and you never stop being in charge.

    1. 01

      Map the work

      I follow how the task is done today, with real examples, and find where agents help and where a person must stay.

      → process map
    2. 02

      Prototype with your data

      A first agent team running on your real cases, so you see the quality before anything is connected.

      → working prototype
    3. 03

      Connect and set limits

      Tools, permissions, budgets and approval rules. The agents start acting on your systems, carefully.

      → connected flow
    4. 04

      Run side by side

      For a while everything goes through a person. We adjust until the approvals are just a click.

      → assisted operation
    5. 05

      Let it run, keep watching

      The team runs on its own. You get a panel with volume, quality and cost, and I keep improving it.

      → monthly report

    06 · Questions

    Frequently asked questions.

    It takes the repetitive part off their plate. People stay where judgment, relationships and decisions matter, and approve what the agents prepare.

    Systems · Web apps · Automations · Websites

    Got a process still living in a spreadsheet? Tell me.

    Fill in the form and the message arrives neatly organized on my WhatsApp. I reply personally, no bots in between.

    What do you need? (pick as many as you like)
    When do you need it?

    Opens your WhatsApp or email with the message ready for you to review. Nothing is stored on this site.

    WhatsApp