Multi-agent System
AI Agents
One service running nine specialised assistants for the CRM and the finance system: a sentence on WhatsApp becomes a categorised expense, a sales call comes back reviewed, and a proposal is drafted and revised.
Some details omitted under NDA
The problem it solved
The same questions came back every week and were answered by hand: was that call any good, who is this company really, where does this expense belong, can you write the proposal. Each answer cost somebody close to an hour, and the quality depended on who did it.
Instead of bolting AI into each system, we built one service of assistants. The CRM and the finance system both call the same service, each with its own request.
What the assistants do
Log an expense by writing a sentence
Send "almoço 45 no nubank" on WhatsApp and a categorised transaction comes back, with a reply confirming what was understood.
A company brief before the call
A lead briefing for the rep to read before dialling.
Deep lead research
The assistant plans its own searches over several rounds, extracts what is new, looks up the Brazilian company registry and enriches from the company Instagram.
Sales call analysis
The conversation comes back as a structured read, to the same standard every time.
Meeting analysis
What was covered, what was agreed and what is still open.
Gatekeeper analysis
A read on whoever answered before the decision maker, one at a time or in a batch.
Call-transfer analysis
When a call is handed to someone else, what happened in the handover.
Proposal writing
Drafting with a revise and correct loop, with the file stored on Drive.
Changes measured, not guessed
Every prompt is versioned, every run is traced, and there are evaluators and experiments to compare before switching.
How it turned out
- Answers in minutes, to the same standard, whoever asked.
- One service serves two different systems, so an improvement to an assistant reaches both at once.
- When we change an assistant's instructions we can compare before and after instead of betting.
- The systems that call the service are not locked to a model vendor, because the AI part lives outside them.
In numbers
9
specialised assistants
2
systems served
4
kinds of conversation analysis
For technical visitors
The rest of this page is written for whoever decides the purchase. This block is for whoever has to maintain the service.
- 9 agent routers across 13,708 lines of Python, with 15 agent test files.
- A standalone service published at agents.wbdigitalsolutions.com, consumed by the CRM and by the finance system.
- Prompt versioning, traced runs, evaluators and experiments with Langfuse.
- Deep lead research plans its own search rounds, extracts updates and enriches from the Brazilian company registry and from Instagram.
- Proposal generation runs a revise and correct loop, with Drive storage.