The problem
Our customers are development teams without a DevOps engineer. When they need something from their infrastructure they write in a support channel and wait for one of ours. We counted every customer request in our own support channels over 60 days: 204 requests from 11 customers. 34% only needed someone with access to look. 37% were changes that could be drafted before an engineer reads them. 28% arrived outside office hours, and a quarter of those waited more than eight hours for a first reply.
What it does
ES On-Call is an agent that works in the customer's ops channel.
- It notices first. When a task of a watched service is stopping, a deployment fails or an alarm goes off, AWS tells the agent. Nobody asks. It investigates and posts the conclusion, the impact, the cause and the state now.
- It turns requests into reviewed changes. A customer writes "add this environment variable" or "scale to 4 tasks". The agent checks the live service, reads the infrastructure code, looks up which inputs that module version accepts in our own module catalogue over MCP, drafts the change as a diff, and hands it to an engineer with a reference.
- It answers from the live environment: what is running, health, deployments, backups, logs, cost.
Why it is safe on production
The agent holds no access of its own. Per customer it assumes a view-only role with explicit denies on secret values, parameter values, object and table contents and decryption. An allowlist and redaction sit in front of that, and Bedrock Guardrails screen what a person types. It has no credential that can change a customer's infrastructure or code: a change is a diff in a message, for a human to apply.
How it uses the AWS AI services
Strands runs the tool loop and streams each step. AgentCore Runtime hosts it with one isolated session per conversation; AgentCore Memory keeps each thread's history per customer. The model on Amazon Bedrock decides what to look at and writes the answer and the diff. Two MCP servers bring in current AWS documentation and the MSP's own platform knowledge.
What exists today, exactly
The channel is a web page backed by a feed, and every post is also sent one way to the engineers' own chat channel. A two-way chat integration, and the engineer's reply returning to the customer, are next and not built. The demo runs one customer environment; the role model and the agent are built for many.
What is new during the hackathon
Everything. The project was started on 13 November 2026.
Built With
- amazon-bedrock
- amazon-bedrock-agentcore
- amazon-dynamodb
- amazon-eventbridge
- aws-iam
- aws-lambda
- bedrock-guardrails
- model-context-protocol
- node.js
- python
- strands-agents
- terraform
