AI can write code in seconds.
It can find a bug, explain what went wrong, and even generate a fix.
But one question kept coming up while we were building:
How do we know that the fix actually works?
That question became the starting point for SentinelLab, the project we’re building at First Commit: Bharat Builds Tour, powered by AWS × WeMakeDevs at Polaris School of Technology, Bangalore.
Along with my teammates, we wanted to build something where AI doesn't just generate a solution and stop there.
The idea
SentinelLab follows a simple flow:
Different agents handle different parts of the process: analysis, security, repair, and verification.
The part we were most interested in was the sandbox.
Instead of simply trusting generated code, we run it in an isolated environment using Amazon Bedrock AgentCore Code Interpreter and use the results as evidence for the final verification.
How it fits together:
We're using Amazon Bedrock, Strands Agents SDK, Bedrock AgentCore, AWS Lambda, API Gateway, S3, DynamoDB, and CloudWatch to bring the whole workflow together.
What fought back?
Quite a lot. 😄
Getting the idea was probably the easiest part.
Making the agents communicate reliably, handling structured outputs, getting AWS permissions right, connecting Lambda with the different services, and getting sandbox execution working all came with their own surprises.
There were also a few moments where every individual component worked perfectly — and somehow the complete system didn't.
That was probably one of the biggest lessons for us: an AI application isn't just about the model. The engineering around it matters just as much.
I also got the opportunity to attend the hands-on workshop “Building AI Agents with Strands SDK: From Simple Loops to Multi-Agent Swarms” by Jatin Mehrotra and Veeramani A.
The workshop gave us a better understanding of agent workflows and how Strands Agents SDK and Amazon Bedrock AgentCore can be used to build more practical agentic systems. Being able to take some of those ideas and apply them directly to SentinelLab made the experience even better.
What we’re taking away
For me, the biggest takeaway is simple:
A model can generate code.
An agent can reason about it.
But trust comes from testing the result.
That’s the idea we’re exploring with SentinelLab.
We're still building, still debugging, and still finding things that break.
But that's also the fun part.
#AWS #AmazonBedrock #StrandsAgents #BedrockAgentCore #AIAgents #FirstCommit #BharatBuilds #Hackathon #WeMakeDevs #BuildInPublic
Loading article