BIOEE 7600-103 · Week 1 · 11 September 2026
Agentic AI in Ecology & Evolutionary Biology
John Benning & Xiangtao Xu · Cornell University · Fall 2026
Public copy of the slides as shown in class, with class-specific detail removed.
Course page: benninglab.org/teaching/agentic-ai
Format
The format
- Wednesday. One of us runs a real task with an agent and posts the whole transcript
to
#demos on Slack. The generation happens before class, not in the room.
- Friday. We walk through the completed task and audit it for quality and
clarity.
- Then we pick next week's prompt from two or three candidates, by vote.
- Nothing to hand in. At most one required reading a week (though we will post lots
of optional readings). Bring a laptop and your questions.
- Seven people offered to show their own five minutes on the survey. We will take
them up on it.
Pre-course survey
A variety of experience with AI
Agentic tools (Claude Code, Cursor, Copilot)
Never used one
10
Tried once or twice
5
Use one regularly
6
Chatbots (ChatGPT, Claude, Gemini)
Most days
7
Most weeks
6
Occasionally or less
8
What you will be driving
At a command line
Comfortable
9
Can run commands I am given
12
Pre-course survey
What you want from an agent
Statistical analysis & figures
14
Literature search & synthesis
12
Data cleaning & wrangling
10
Simulation modeling
9
Writing (manuscripts, grants)
8
Genomic / bioinformatic pipelines
5
Lab & project management
4
Teaching material
1
What is wrong with this figure?
Pre-course survey
What you want from an agent
Statistical analysis & figures
14
Literature search & synthesis
12
Data cleaning & wrangling
10
Simulation modeling
9
Writing (manuscripts, grants)
8
Genomic / bioinformatic pipelines
5
Lab & project management
4
Teaching material
1
Not a hypotheticalThe previous slide is a real bug from an earlier build
of these slides, put back in on purpose. The numbers were right, one line of CSS was wrong, and
the chart still looked like a chart. Visual acuity is one weak spot for these models,
currently.
Pre-course survey · recurring themes
What worries you
That efficiency is not what several steps of the research process are for, and that wide
adoption will bring pressure to do everything faster anyway.
Accountability. Whether you will be able to check and validate what comes back.
Running analyses you do not understand, and code you could not read.
Overuse in writing, overreliance in analysis, and people who believe an agent can stand in
for a student.
Losing your own skill at the thing. And the energy it costs.
Paraphrased, not quotedThese are themes drawn from free-text survey
answers, rewritten so no one is quoted verbatim. Also raised: reproducibility · replacing humans in research
decision-making · privacy of sensitive data · accuracy in less-studied systems
· what the companies do with what we type · unclear expectations about disclosure
· how to use it in writing without losing your own voice. Week 8 is energy, run on our own
logged numbers. Week 9 the class drafts an EEB guideline.
How John uses aAI (agentic AI) · 1 of 3
My basic setup
How John uses aAI · 2 of 3
Three example projects
Fringed gentian · coordination
- Range-wide leaf collection by volunteers — naturalists, preserve managers, agency
botanists. 56 populations, 20 states and provinces, no field crew, a three-week flowering
window per site.
- 687 emails across 195 threads, 317 of them sent by us.
- The agent keeps the tracker, the permit ledger by jurisdiction, and helps draft emails.
Nothing is sent without human eyes on it.
Finger Lakes vegetation · exploratory analysis
- 181 plots surveyed around 1980. The data sheets survived; the location maps did
not.
- The agent wrote the geospatial pipeline that turns one line of text — quadrangle,
slope, aspect, elevation — into a candidate polygon on a map.
- It also found sampling confounds and small data errors.
Clarkia crosstypes · making analyses more robust
- Building robustness checks for a model whose structure is genuinely labile.
- Power simulations that let us say more concretely whether an absence of an effect is
likely real or a power issue.
- An audit of the processed data handed over from others against the raw
sheets, all 19,956 rows: three disagreements, one substantive.
How John uses aAI · 3 of 3
Some general insights
- Use aAI to automate tedious tasks.
- Build in lots of checks, internal and external.
- Exploratory analyses outside your realm of expertise are OK, but don't fool yourself
into complacency.
- HTML reports are great.
- Use aAI to do better science, not necessarily more science, and to learn more, not
less.
- Persistent memory changes everything.
Fifteen minutes · small groups
Small group discussions
- Hands up if you have used an agentic tool — Claude Code, Cursor, Copilot,
Codex — more than once or twice.
- Everyone else, sort yourselves around a raised hand. Three or four to a group, one
raised hand each if it works out.
Prompts
- What is the most tedious recurring thing in your week that you think aAI could possibly
help with?
- What have you tried handing to AI, what has it done well, what has it done poorly?
- What's the best case and worst case scenario for aAI in EEB, in your mind?
For next week · agentic coding and data analysis
What should we run?
Non-linear gradients
Recent theory on species range limits suggests that non-linear environmental gradients
may be especially important for limiting distributions.
Skeleton prompt
For 20 well-sampled North American plant species, pull occurrence records from GBIF and
climate from WorldClim. For each species define the range core and the range margin, and
compare the local spatial rate of change in a set of putatively important climatic variables
between them. Give me robustness checks and do not chase results.
Reproduce a published result
An option with an answer key. The archive ships the SLiM model, the parameter files, the
R code and the simulation output the paper was built from.
Skeleton prompt
The archive for Benning et al. 2024 (Ecol. Lett. 27:e14472) is public at
doi:10.6084/m9.figshare.c.6981030. Run the model at the paper's focal parameters and
reproduce the front-versus-core difference in dispersal phenotype on a flat landscape. Do
not quit until you can exactly reproduce the paper's results.
A dataset with something wrong in it
A file built on one of my own field experiments — 160 mesocosms at a
Clarkia range edge, soil source crossed with precipitation (Benning & Moeller
2021, Ecology). We plant faults in it and do not say what. The published result is
the answer key.
Skeleton prompt
Attached is a cleaned dataset from a factorial field experiment, and its data dictionary.
Does soil environment affect lifetime fitness? Fit an appropriate model and tell me whether
the effect is significant.
Other?Let's brainstorm. What would you like to see re: aAI coding and data analysis? The transcript goes up in #demos on
Wednesday and you can re-run it, alter it, argue with it. None of it is homework.
Before next Friday
Two things, neither of them homework
- Join the Slack.
#setup-help for install trouble,
#demos for what an agent actually did, #general for everything else.
Announcements go there and to the course page, not to your inbox.
- Use your Claude for Science invitation if you need an account. Access is covered for
everyone in this room — nobody should be paying for this.
Optional, if you want itThe "what is an agent?" primer on the course page
walks a real GBIF example from chatbot to agent. Entirely optional — you will not be behind
if you skip it.