TL;DR
Our client, an early-stage startup in Bengaluru, is hiring engineers to build agent systems that run in production rather than in demos. The distinction is the whole job: agents that work once in a notebook are common, agents that behave predictably enough to put in front of a customer are rare.
The client is confidential at this stage. We will tell you who it is before your profile goes anywhere.
What the work is
Designing and building systems where a model plans, calls tools, and acts over multiple steps. That means the engineering around the model more than the model itself: tool interfaces that fail safely, state and memory that survive a step going wrong, retries that do not compound errors, and evaluation that can tell a genuine improvement from a lucky run.
Observability matters more here than in most engineering. When an agent takes eleven steps and the ninth is wrong, you need to be able to see why, and most teams discover too late that they cannot.
What the client is looking for
Engineers who have built something agentic and can describe its failure modes honestly. The best signal is someone who explains what their agent got wrong, what they changed, and how they knew it worked, rather than someone who describes only the happy path.
Strong software engineering underneath. This is a systems role with a model in it, not a prompting role, and the people who do well are engineers first.
Comfort with early-stage ambiguity. The product is still being found, so the brief will change, and candidates who need a settled roadmap will not enjoy it.
Process
A conversation with us, then the client's founders, then a practical session building or debugging something agentic.
Related reading
Agentic AI in trading and agentic AI in quant research are worked examples of the same engineering problem under harder constraints.
Interested in this role?
Apply once, to us. We confirm the client with you before anything is submitted. Because of application volume, only shortlisted candidates are contacted.