Um plano de contingência que avisa sozinho: MCP, agent e WhatsApp no Amazon Quick

Blumenau, a city located on the banks of the Itajaí-Açu River, is constantly dealing with flood levels. Dati maintains a Flood Contingency Plan which defines, for each flood level, the actions each team must take. The plan has existed for a while, but lacked a real-time indicator of the current flood level. On September 11, 2026, the Defense Civil department projected a water level of 9 meters for 2 AM, prompting a WhatsApp auto-response to leadership.

Setting up this automated delivery took only 27 minutes. El Niño returned, bringing back the same question. The Santa Catarina Civil Defense confirmed El Niño in June 2026, estimating a 90% chance of its formation during winter, with strong to very strong intensity expected from spring to summer, and noting an increase in effects starting from September.

The Itajaí Valley is highlighted as the region historically most affected. Those working in Blumenau are familiar with the subject without needing technical notes. A prolonged forecast of rainfall reorganizes everyone's routine, from schools to businesses. Dati's plan outlines actions for each river flood level, from initial notification to leadership activation.

Such a plan depends on a single piece of information to function: the current river level. Without this number, no level can be triggered. The number exists, and its consultation was previously done manually. A Civil and Defense Protection agency in Blumenau publishes the river level on an open panel, powered by the AlertaBLU system.

The panel's left side defines the levels, while the right side shows the current measurement. Each level corresponds to a set of actions from Dati's plan. The "Alert" level, starting at 6 meters, triggers the continuity continuity plan, monitoring news channels and river level, and activates leadership in a risk situation. Operational guidance follows these three lines: removing laptops at the end of shifts, ensuring mobile phones and power banks are charged, and preparing homes, families, and emergency kits for those living in risk areas.

Previous routines had three tasks. Someone would open the panel, find the most recent measurement, compare it to the table, and then alert the team. The current update only once per hour. A flood lasts days. Multiplying one by another results in dozens of manual queries, hence the subject becoming an engineering task. July: the space, the agent and the MCP server The complete path.

The MCP server runs on Dati's AWS account, while the agent and flow live within Amazon Quick. On July 3, 2026, the Flood Contingency Plan moved into an Amazon Quick space - the library of documents other product functions query - as a single knowledge document, with slightly more than two megabytes. Any Quick resource consulting that space sees the entire plan.

The content stopped depending on who remembers where the file was saved. On July 6, the Dati Flood Contingency Agent was born, supported by that space. Its purpose is to answer employees' questions about safety procedures, flood alert levels, and actions in flood risk situations. The screen offers three ready-made questions: current river level, evacuation meeting points, and procedure for the Observation level.

It was built conversationally, so a person can ask and the agent will answer. The instructions determine four behaviors: a closed scope. Responses are limited to the plan's content, explicitly prohibiting assumptions about information absent from the document. Outside scope, the agent recognizes the absence and recommends contacting leadership.

Fixed ranges. Each level's numerical limits, with corresponding colors, are written in the instructions. Classification is no longer a model's judgment. Live. The instruction explicitly discourages reliance on knowledge-grafted trust. Unlike the plan stored in the space, which lacks current measurements, this agent's live response would provide an outdated number.

Mandatory format of output. Level, trend, classification, and update time, followed by a descriptive paragraph and a conclusion about the need for action. A free-form agent delivers different text each execution, making variable-output-based automation fragile. On July 8, a third component entered the picture: an MCP server exposing the river level as a tool.

The server ran on Dati's AWS account and consumed the municipal Defense Civil public feed. The feed returned the last 24 hourly readings, with level and time for each, and the table of levels with each level's floor in meters. Choosing a MCP server over the agent fetching the page alone had three reasons: Contract. A tool declares name, description, and parameters, so the model knows the information exists and how to request it.

Reuse. The same server serves the agent, flow, and any MCP-speaking client without duplicating collection code. Authentication type. Amazon Quick authenticates an MCP connector via client credentials, a service authentication. This property has consequences later. The server doesn't transform the data. The tool returns the response from the source as it arrives, so the consuming model finds the most recent reading and crosses the value with the table on its own.

This decision was deliberate. By end of July, knowledge, classification, and collection were resolved. Distribution remained manual. September: the WhatsApp connector June 2026 saw the addition of a WhatsApp connector to Amazon Quick. Dati Labs opened a ticket about its viability in early September, with a question from another project: the feasibility of using audio received via WhatsApp.