Hi, I'm Tushar,
an AI Research Engineer.
I build the infrastructure for autonomous AI — multi-agent runtimes, evaluation systems, and the coordination layer that makes agents reliable. Currently at Atomicwork.
About
I'm an AI Research Engineer at Atomicwork, where I built the multi-agent runtime powering our agentic service management product — a Gateway-based harness with per-agent context isolation, MCP server integrations, YAML workflow orchestration, and subagent spawning.
The core problems I live in: stateful multi-agent coordination, domain-specific agent evaluation, and code agent harness design. Before Atomicwork, I was at SuperAGI Research building autonomous coding agents (SuperCoder 2.0 — #1 open-source on SWE-Bench), cognitive GUI automation (AUTONODE), and multi-agent orchestration frameworks.
I studied ECE at Netaji Subhas University of Technology (Delhi), graduating with dual minors in ML and Signal Processing, ranked 3rd in my department. I've also done research stints at ISRO and NTNU Norway.
Outside of work, I write — poetry, prose, and the occasional piece on things that aren't code.
Currently Working On
Building Synapse — a unified agent runtime with online evaluation pipelines and built-in observability. Also exploring RL-enforced recommendations for a stealth project.
Highlights
Atom Agent Harness
Multi-agent runtime with per-agent context isolation, MCP servers, YAML workflow orchestration. 55% latency reduction, 3x throughput.
SuperCoder 2.0
#1 open-source on SWE-Bench. Autonomous software development with 34% on SWE-Bench Lite, Top 3 globally.
AUTONODE
Cognitive GUI automation for self-operating web navigation. Accepted at ICAC 2024.
Papers
Contact
Interested in collaborating on multi-agent systems, agent evaluation, or runtime infrastructure? Let's brew some conversations over coffee.
Get in touch →