Speaker: Aaron Tay, Academic Librarian, Singapore Management University
Title: Agentic AI might be changing everything again
Subtitle: What an LLM with tools can actually do for information work, and things to be cautious about
*This webinar will not be recorded
Description:
Just as librarians were getting comfortable with chatbots, the ground shifted again. Agents or LLMs that use tools in a loop - can now search, read what they find, reformulate, retry and recover from dead ends the way a human expert does. Early adopters are already building powerful flows on harnesses like Claude Code and Codex, which marry the flexible but non-deterministic nature of LLMs to the determinism of code. In academic search specifically, agentic search has arrived.
The landscape is moving on two fronts. Search startups such as Undermind, Elicit and Consensus now claim to support more agentic search on their platforms. At the same time, a widening set of providers is shipping MCP (Model Context Protocol) servers that pair with an LLM to allow building of powerful, sophisticated home-brew agentic skills: Wiley, Scite, Consensus and Elicit are live (together with unofficial MCPs to free search tools like PubMed, OpenAlex, Semantic Scholar etc), with EBSCO and Clarivate announced.
In this talk I will give some clarity on how agentic search actually works, how it relates to "deep research" and "deep search", and what the existing empirical evidence says about how well it performs on modern tough information-retrieval benchmarks like BrowseComp-Plus.
Closer to home, I have found that linking even a relatively weak LLM to a simple, home-brew Primo MCP server resulting in a LLM that can iterate on its own prior results - produces surprisingly large improvements in database discovery for inexperienced users, well beyond simple synonym expansion.
I will briefly highlight two further examples: an agentic skill that uses PubMed and MeSH tools to construct and pilot highly sensitive PubMed search strategies for systematic reviews, and an advanced lit-review orchestrator that combines Undermind, SSRN and other avenues with deduplication and validation against hallucination.
But putting an LLM in the search loop is already known to hurt the interpretability and reproducibility of results. Does agentic search make that worse? I will share partial findings from my own mini-study on exactly that question.