In the ever-evolving landscape of artificial intelligence, the development of Deep Research Agentic Systems has emerged as a groundbreaking innovation. These systems, exemplified by OpenAI and Gemini Deep Research Agent, are designed to conduct multi-step research on the internet, employing dynamic reasoning and multi-hop information retrieval. At the Arc of AI Conference 2026, Sarang Kulkarni, a Thoughtworks team member, shared invaluable insights into the design and deployment of multi-agent research systems for deep reasoning and synthesis, drawing from real-world healthcare and pharmaceutical R&D projects. Kulkarni's presentation shed light on the challenges and lessons learned in building these sophisticated AI agents.
One of the critical industries that benefit from these advancements is healthcare, where researchers require more than traditional AI models. They need systems capable of discovering, connecting, and reasoning across both internal and internet data while ensuring reliability, transparency, and compliance. Kulkarni emphasized the staggering cost of bringing a new drug to market, estimated at $2.6 billion, and highlighted the issue of half the research studies being conducted without prior evidence due to broken knowledge access.
To address these challenges, Kulkarni's team developed a Retrieval Augmented Generation (RAG) based chatbot two years ago to search through unstructured data. While the RAG solution proved effective for simple queries, complex questions demanded enhancements, leading to the creation of the Agentic RAG++ application. This system comprises a clarification loop, research loop, and writing loop, each serving distinct purposes in the research process.
The initial version of the researcher agent was built using two tools: the RAG tool and the text2sql tool. The RAG tool employs weighted hybrid search, 20 context chunks, a re-ranker, and seven refined context chunks to facilitate information retrieval. Meanwhile, the text2sql tool plays a crucial role in improving query execution accuracy by feeding SQL query errors back to the LLM. However, Kulkarni warned that factors like higher token costs, poor performance, and high latency can hinder effective retrieval from AI agents.
Context anxiety, as Kulkarni noted, is another challenge that teams must navigate. Incomplete data can lead to poor self-evaluation, but the reflection loop technique can help address data completeness. Kulkarni discussed the various failure modes encountered during the development of the custom deep research agent solution, emphasizing the importance of an explicit think-act loop for long-horizon tasks.
Long-horizon tasks often break decisions between steps in the process. To overcome this, Kulkarni's solution incorporates a reflection step that includes data reflection and process reflection. This phase also introduces a Draft Writing Loop to address synthesis gaps, ensuring that all relevant information is captured in the final report. Kulkarni concluded by highlighting the emergence of harness engineering techniques, which focus on designing tools, memory systems, and validation checks to make autonomous AI agents more reliable and accountable.
Harness engineering aims to shift AI solutions from prompt engineering to automated task execution by AI agents. Kulkarni emphasized that the better the models, the thinner the harness needs to be. This approach transforms AI agents into a combination of models and harnesses, ensuring their effectiveness and reliability in complex research tasks.
