Insights from Sarang Kulkarni: Key Lessons Learned in Developing Deep Research Agents for Production

### Understanding Deep Research Agentic Systems

Deep Research Agentic Systems, such as OpenAI and Gemini Deep Research Agent, represent a groundbreaking leap in artificial intelligence technology. Specifically designed for conducting multi-step research on the internet for complex tasks, these AI agents utilize dynamic reasoning and multi-hop information retrieval. Their primary strength lies in their ability to generate comprehensive, structured analytical reports akin to those produced by seasoned research analysts.

### Insights from the Arc of AI Conference 2026

At the Arc of AI Conference 2026, Sarang Kulkarni from the Thoughtworks team shared invaluable insights into the design and deployment of these advanced multi-agent research systems. His talk focused on the principles of deep reasoning and synthesis, derived from practical applications in healthcare and pharmaceutical R&D projects. Kulkarni emphasized the innovative techniques his team employed, including agentic loops and harness engineering, to optimize their solution’s performance.

### The Need for Advanced AI in Healthcare and R&D

In critical industries like healthcare and clinical trials, traditional AI models that provide simple Q&A capabilities are insufficient. Researchers require systems capable of discovering, connecting, and reasoning across diverse data sources, including internal databases and the vast expanse of the Internet. Such systems must not only maintain reliability and transparency but also comply with stringent industry standards.

### The High Cost of Drug Development

Kulkarni began his presentation by highlighting a stark reality: it typically costs around $2.6 billion to bring a new drug to market. Remarkably, nearly half of all research studies are conducted without prior evidence, often due to a lack of access to existing knowledge. This challenge is particularly evident in the drug discovery pipeline, where obtaining the right data at the right time proves to be a daunting task.

### Creating the Agentic RAG++ System

To address these challenges, Kulkarni’s team developed a Retrieval Augmented Generation (RAG) based chatbot two years ago, designed to sift through unstructured data. While this initial RAG solution worked well for simple queries, it proved inadequate for complex inquiries. Consequently, they evolved their approach into an enhanced system called Agentic RAG++.

This sophisticated framework comprises several crucial components. One key feature is the clarification loop, which helps to ensure that queries are well-defined. The research loop includes tasks like thinking, planning, executing, reflecting, and adjusting plans based on findings. Lastly, the writing loop focuses on documenting insights and reflecting on those documents for clarity and comprehensiveness.

### Tools for Enhanced Research Capabilities

The initial version of the researcher agent utilized two main tools: the RAG tool and a text2sql tool. The RAG tool, designed for weighted hybrid searches, processes multiple context chunks to enhance retrieval accuracy. Meanwhile, the text2sql tool aids in feeding back SQL query errors to the AI model, thus improving the accuracy of query execution over time.

### Challenges in AI Retrieval and Solutions

Kulkarni pointed out several potential pitfalls that researchers must navigate when using AI agents. These include high token costs, poor performance, and latency issues that can stem from incomplete data. To combat these challenges, the reflection loop is employed, which helps verify the completeness of data while also prompting the improvement of processes.

### Addressing Long-Horizon Tasks

When it comes to long-horizon tasks—those that require multiple steps to complete—Kulkarni emphasized the necessity of having an explicit think-act loop in place. Incorporating multiple steps like think, plan, inspect, and update ensures that the research remains coherent and produces useful outcomes. For this, they utilized solutions like Anthropic’s “think” tool, which provides a structured pause for reasoning.

### The Importance of Reflection Steps

The reflection phase in Kulkarni’s model extends beyond simple data verification; it encompasses a process reflection that assesses the overall completion of the project. An additional third step, termed the Draft Writing Loop, addresses any synthesis gaps. This ensures that no pivotal information derived from research is lost during the writing process.

### The Future of Harness Engineering in AI

Kulkarni concluded his presentation by discussing the emerging field of harness engineering. This approach focuses on designing tools, memory systems, and validation checks that make autonomous AI agents more reliable and accountable. The goal of harness engineering is to elevate AI solutions from mere prompt engineering to fully automated task execution.

In essence, since AI agents merge models with harnesses, the quality of the models directly influences the required complexity of the harness. The better the models are, the simpler the harness can be, paving the way for more efficient and effective AI systems in deep research applications.