Agentic AI · AWS · 2026

An AI pipeline on ECS, driven by a custom MCP server

Professional services client

Python agents running as containerised tasks on ECS Fargate, invoked through a purpose-built MCP server so the client's own tools could call them like any other function.

24/7Agent availability, no human in the loop
<2sTypical end-to-end invocation
100%Invocations logged and auditable

The problem

The client had proven that an AI workflow could save their team hours a week, but it lived in a notebook on one person's laptop. Nothing else in the business could reach it, nobody could audit what it had done, and there was no path to running it reliably.

What we did

We packaged the agents as containers and ran them as ECS Fargate tasks, then put a purpose-built Model Context Protocol server in front so the client's existing tools could invoke an agent the same way they would call any internal API. IAM was scoped per agent rather than per cluster, CloudWatch captured every invocation for audit, and deployment moved to a CI/CD path from GitHub through CodeDeploy into Fargate.

The same pattern now backs the appointment-booking agents we embed in customer websites.

Scoping a project? Start here.