Skip to content

Agent agency

AI agents we use before we sell them.

Xispe designs, integrates and runs agents for companies. The same way of working powers our own product, De Olho no Fundo.

The path of a document in De Olho no Fundo
  1. fnet-sync

    Finds the new document on FNET, B3’s official filing system.

  2. extract-pdf | extract-xml

    Turns the PDF into text and the XML into fields.

  3. analyze-document

    Summarizes what changed for the unitholder, with impact and relevance.

    Off format or sounds like advice: it fails and retries.

  4. notify-document

    Alerts everyone following the fund, with a link to the original.

What we do

Custom agents

We take a repetitive job that eats up your team’s time, like reading documents, triaging requests or drafting replies, and build the agent that does it and hands back a result your systems can use.

In De Olho no Fundo Every document becomes validated JSON, with a summary, the direction of the impact and a relevance score from 1 to 5.

Integration

We connect the agent to your company’s data and tools. With MCP, the open protocol that gives AI assistants access to tools, your system can also serve your customers’ own agents.

In De Olho no Fundo An MCP server with 11 tools, from reading a document to setting up alerts, with API tokens and read or write permissions.

Operations

An agent needs an owner after launch. We track cost, failures and quality, and tune the prompt and the model based on what the numbers show.

In De Olho no Fundo A monthly spending cap, fallback models, Discord alerts and a quality report by pipeline stage and prompt version.

Our own product

De Olho no Fundo

Our product for people who invest in Brazilian real estate and agribusiness funds. A portfolio of 15 funds publishes around 50 documents a month on FNET, the official filing system of B3, the Brazilian stock exchange, and a management report runs 26 pages on average. The agents read all of it and tell unitholders what changed.

Pre-launch

Visit De Olho no Fundo

The product is for Brazilian investors, so its site is in Portuguese.

Agents inside the product

An agent reads every document

Management report, material fact or dividend notice: it reads the whole document and writes, in a few lines, what changed for the unitholder.

The customer’s agent joins in

Through the MCP server, Claude Code, Cursor and other assistants our users already have search the documents and compare dividends right in the conversation.

Agents building the product

The API, the pipeline, the web app, the Android app and the product’s own website were built pairing with coding agents. So was this site.

87 of 93
product commits made pairing with Claude Code
61
automated test files
11
tools for agents on the MCP server

Figures as of November 3, 2026.

De Olho no Fundo does not give investment advice: it summarizes what the funds published and keeps the original document one click away.

Rules our product taught us, applied to your project

  • The numbers don’t come from the model

    Dividends, record dates and net assets are read straight from the XML filing and override whatever the model wrote.

  • Off-spec output doesn’t get through

    Every response is validated. If it is off format, or reads like investment advice, the step fails and retries.

  • The source is one click away

    Agents can be wrong, and we tell our users so. Every summary links to the original document.

  • Cost has a ceiling

    Model spend is measured per analysis and capped every month. The agent only runs where someone will read the result.

  • Switching models takes no deploy

    Models get retired and prices change. The main and fallback models live in configuration, and comparing them takes a single command.

  • Quality is measured

    Every analysis can be voted on, and the votes are broken down by prompt version and by model.

Got a job that sounds like agent work?

Tell us what the job is, who does it today and how long it takes. We’ll tell you whether an agent can handle it, even when the answer is no.

[email protected]