Workshop for researchers in all disciplines
Agentic AI in the Scientific Research Workflow
Accelerating the speed of discovery, without the surprises.
- Date
- Tuesday, September 1, 2026
- Time
- 10:00 am – 12:30 pm
- Location
- PGH 563, University of Houston
- Hosts
- Larry Shi and Omprakash Gnawali, Computer Science
The research gain
Give an agent a data-analysis task and it writes the code, reshapes the data, makes the figures, and drafts the interpretation, chaining the steps at record speed. The tedious middle of research shrinks from days to minutes.
The OECD argues that raising research productivity may be AI’s most valuable application. And the gain is larger than benchmarks suggest: in science, the unit of performance is the scientist and the agent working as a team.
The catch
The agent makes calls you never see, and reports none of them. Nothing announces the error. The analysis runs; the numbers look right.
It surfaces late, at replication, at review, or in someone else’s reanalysis, with your name on it.
You do not have to choose between the two columns. That is the point of the workshop.
Is this workshop for you?
- Your group’s work is data-heavy and quantitative.
- Someone in your group has started using AI to write analysis code, or will within a year.
- You are accountable for results you did not personally produce.
- You’ve heard about AI agents and have questions you haven’t had anywhere to ask.
Two or more? This session is aimed squarely at you, including if you have never used an agent yourself. No AI expertise is assumed; workflow is the common language. PIs, postdocs, research-computing and core-facility staff from chemistry, biology, economics, pharmacy, materials, biostatistics, data science, health sciences and their neighbors.
What we cover
The focus is the data-analysis and processing workflow. These failures come from the agents, not from your science: a chemist, an economist, and a biostatistician hit the same ones. That is why the room is mixed.
- The code you never asked for
- You asked a research question; the agent produced code. It runs once, is rarely reviewed, often isn’t kept, and it decides whether your result is right.
- Data an agent can actually read
- Where the meaning of data is left implicit, the agent supplies a plausible one. It may not be yours, and nothing flags the substitution.
- Can you retrace your steps?
- Six months on, a reviewer asks how a number was produced. Which data version, which settings, which code? Right but unreconstructable is still a problem.
- What you didn’t get told
- Agents answer the question asked. They don’t volunteer that a rival explanation fits just as well, so real findings get missed quietly. We call this re-seeing the data.
- Working with a non-human colleague
- Which judgments must stay with you; what happens to a trainee who learns the work by prompting; how oversight slips one reasonable delegation at a time.
How the morning runs
Half of it is yours. Places are kept small so the discussion stays a discussion.
We present (60 min)
- 50 min
- Agentic AI workflow in data-oriented sciences; agents as research assistants; the failure modes and why they stay invisible; a reusable draft protocol; human–agent teaming in research tasks.
- 10 min
- Open Q&A with the presenters.
You take part (60 min)
- 20 min
- Five-minute talks from attendees: what your group uses agents for, and one thing that surprised you.
- 25 min
- Small mixed-discipline tables.
- 15 min
- Synthesis across the room.
What you walk away with
- A field cardA one-page summary of common failure modes and the checks that catch them. Usable the next morning.
- A working protocolWhat to track, what steps to follow, a rule-book and best practices: yours to argue with and test in your own group.
- Straight answersTo the questions you’ve been collecting and had nowhere to put.
- Colleagues across fieldsFacing the same decision from a different direction.
Attendees
The list of participating groups and disciplines will be posted here after the workshop.
- Name — Department
Outcomes
Materials from the workshop will appear here as they are finalized: the field card, the draft protocol, and a synthesis of the discussion across disciplines.
Most of this material isn’t in the literature yet. It comes from research teams building and testing agentic AI platforms for science, documenting the failure modes, tracing the damage, and building defenses for productive science in the era of AI-native discovery.
Beyond the workshop
We are open to feedback, follow-up discussion, and suggestions on agentic AI for the sciences. Nothing to decide on the day. The door is simply open, always.
Hosted by Larry Shi and Omprakash Gnawali, faculty of Computer Science, University of Houston.