Integrating Tools and Orchestrating Agents for Krutrim's Assistant

Karan's scope covered the infrastructure around Krutrim's assistant agents. He maintained the available tools and the engineering that connected them to the agent system, supporting reliable tool selection and efficient task completion.

He also built the multi-agent orchestration layer. The assistant had six to eight specialized agents, each responsible for a distinct task domain such as food ordering, cab booking, bill payments, and map-based ride search. When a user query involved multiple domains, Karan's orchestration logic broke the request into sub-queries and routed each to the appropriate agent.

During production operation, a scheduling agent developed a runaway behavior: a developer-introduced bug allowed the agent to schedule itself recursively, creating an unbounded loop of tasks. Karan diagnosed the issue and fixed it by adding hard limits and blocking the agent from accepting recursive scheduling instructions.

From the results he recalls, tool-call success was approximately 95 to 96%, though he treats that figure as directional rather than a formally published measurement.