Beyond the Prompt: Building Research Agents in Python
Speaker
Simona Skiotytė
As a Software Developer for the Generative AI platform at Nasdaq, Simona Skiotytė specializes in architecting the integration of Generative AI systems on AWS. Key projects include the integration of a research system that automates query generation, iterative web search, and multi-stage report synthesis, all orchestrated in AWS. Simona's experience extends to building secure and scalable cloud infrastructure. She architected a multi-tenant data platform with comprehensive test coverage and audit logging. Simona also developed a framework for an AWS cross-account data ingestion into vectorstore featuring complex IAM roles, presigned URLs for secure file management with KMS encryption, and lifecycle orchestration using AWS Step Functions. Additionally, she built a CRUD API for managing version-controlled, reusable Generative AI prompt templates. Simona's foundation in data science comes from her time at ETH Zürich, where she applied statistical and mathematical models to large-scale biological datasets, utilizing dimensionality reduction, clustering, and regression modeling. This background provides a strong analytical lens through which Simona approaches the construction of Generative AI systems.
Abstract
LLMs can answer questions, but can they conduct research? Real research requires planning, data gathering, synthesis, and refinement. This talk explores Python patterns for orchestrating multi-step workflow to create autonomous research agents. We'll deconstruct the agentic architecture, from DAG-based planning to reflection loops, and show how to implement these concepts using practical Python patterns.
Description
While early LLM API calls were stateless and frozen in time, the landscape has evolved dramatically with the rise of "thinking models." OpenAI’s o1, Gemini's Thinking models introduced reasoning first architecture, where the models stops, reflects and elaborates on their thinking, and produces output that follows logical steps; Furthermore, Gemini integrating Google Search which leverages its searching capabilities, make the models not only access the vast information available, but reason and decompose complex problems in real-time. To harness this power and move from simple Q&A to true deep research, we must shift our approach from mere prompting to building the autonomous, multi-step, chained agents that these new models enable. The session will cover:
- Agentic Shift - moving beyond prompts to building agents, and using Pydantic to force LLMs into structured "thought processes". We'll explore LLM tools, memory and ability to adapt based on environment feedback.
- Decomposing the research process into specialised agent nodes: Planning Agent that decomposes queries into a DAG of sub-tasks, parallel Executor Agents that search, retrieve, and evaluate sources, and a Synthesizer Agent that aggregates findings into a coherent report.
- Reflection & Quality Control – Implementing gap analysis where the system reads its own draft, identifies missing citations or thin sections, and formulates follow-up queries. We'll also cover the LLM-as-Judge pattern—using an LLM to evaluate retrieved chunk relevance, weigh source authority (preferring a 2024 paper over a 2022 blog post), and discard noise before synthesis.