An overview of AI agents
Andrew White
Edison Scientific
Regeneron
12 October 2026
A Brief History of AI
What is AI?
Two kinds of machine learning
Generative
Learns from data: examples of \(x\)
Models \(p(x)\) and samples new \(x\)
Example: propose a new molecule
Regression / Classification
Learns from labels: pairs \((x, y)\)
Models \(y = f(x)\) and predicts \(y\)
Example: predict microsomal stability
Breakthrough classification: AlexNet 2012
\(x\): image
\(y\): label
- leopard
- jaguar
- cheetah
- snow leopard
- Egyptian cat
1 of 1,000 classes
Top-5 error: 15.3%
(runner-up: 26.2%)
SSI
Calling for slowdown of AI progress
MIA
“An unequivocal turning point in the history of computer vision” — Yann LeCun, ECCV 2012
Krizhevsky, Sutskever & Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NeurIPS 2012. Photos: tropicaLiving / Wikimedia Commons (CC BY-SA 3.0); Johnny Guatto / University of Toronto (2013)
Graph neural networks for molecules
Molecule = graph. Atoms pass messages to their bonded neighbors.
Permutation equivariance: renumber the atoms and the prediction doesn't change.
Neural fingerprints
Learned convolutions on molecular graphs
Duvenaud, …, Aspuru-Guzik, Adams
GCN & MPNN
General message passing on graphs
Kipf & Welling; Gilmer et al. (Google)
Chemistry predictions
Quantum-chemistry properties at chemical accuracy, in ms
Gilmer et al.: 11 of 13 QM9 targets
AlphaFold2
Equivariant attention over residues predicts protein structure
Jumper et al. (DeepMind)
Duvenaud et al., Convolutional Networks on Graphs for Learning Molecular Fingerprints, NeurIPS 2015; Kipf & Welling, Semi-Supervised Classification with Graph Convolutional Networks, ICLR 2017; Gilmer et al., Neural Message Passing for Quantum Chemistry, ICML 2017; Jumper et al., Highly accurate protein structure prediction with AlphaFold, Nature 2021. Animation: White, Deep Learning for Molecules & Materials, dmol.pub
Transformers: 2017
Encoder reads the data \(x\)
The cat sat on the mat
Decoder writes the label \(y\), one word at a time
Die Katze saß auf der Matte
Attention connects every word of \(y\) to every word of \(x\)
When \(y\) is a sequence, prediction becomes generation
Vaswani et al., Attention Is All You Need, NeurIPS 2017. Figure reproduced with permission for scholarly use.
Decoder-only generative model
Radford & Sutskever (OpenAI, 2018): decoding is enough
Large language model (LLM): a decoder-only transformer trained to predict the next word across huge amounts of text
Radford, Narasimhan, Salimans & Sutskever, Improving Language Understanding by Generative Pre-Training, OpenAI 2018. Detective-novel example: Sutskever, GTC fireside chat with Jensen Huang, 2023.
From GPT-2 to o1
GPT-2
Scaling up data and compute improves performance
GPT-3
Generalizes to new tasks from just an instruction
GPT-4
Scale brings the ability to write code and use tools
o1
Can change how long it thinks before answering
Radford et al., Language Models are Unsupervised Multitask Learners, 2019; Brown et al., Language Models are Few-Shot Learners, NeurIPS 2020; OpenAI, GPT-4 Technical Report, 2023; OpenAI, Learning to Reason with LLMs, 2024
What is AI?
AI Progress
Model intelligence doubles every ~4 months
METR Task Completion Benchmark metr.org
Effect is Visible in Economy
AI models are superhuman in math
Millennium Prize
OpenAI reports a solution to Navier–Stokes in September 2026.
Only 1 other Millennium Prize problem solved in 25 years.
ARC-AGI
2025: o3 scores 87.5% for ~$500,000.
2026: Astra scores higher for ~$20.
Math Olympiad
OAI, Ant, and many start-ups got perfect scores 42/42 on IMO 2026.
1 Millennium Prize / NYT 2 ARC-AGI / Noam Brown 3 IMO / China Daily
1 NASA OIG Report oig.nasa.gov 2 US FHWA fhwa.dot.gov 3 US Telecom Capex Report ustelecom.org 4 Morgan Stanley AI Market Trends 2026
What about in science?
Labbench 2
1,912 questions across accessing data, reading patents, reading figures, protocol understanding, etc.
Example: Deletion of which residues from C. elegans protein COSA-1 would most likely affect the ability of COSA-1 to recruit MSH5 and ZHP3?
Residues 31–40
Labbench 2 Accuracy
FutureHouse
- FutureHouse founded in 2023
- Andrew White, Sam Rodriques, Eric Schmidt
- Based in San Francisco with Wetlab
- Raised $35M
FutureHouse Timeline
Edison Scientific
- Spinout from FutureHouse formed in 2025
- AIxBio Research Lab that partners with biotech and pharma
- 50 employees
- $70M in seed financing
Introduction to AI Agents
Agent: trained, makes decisions
Environment: untrained, has tools, state
How are agents trained?
- Pre-training — broad knowledge via next-token prediction
- Supervised Fine-Tuning (SFT) — learn from expert instruction–response pairs
- Reinforcement Learning (RL) — optimize through environment interaction with verifiable rewards
RL expands capabilities beyond what supervised data can teach
Verifiable Rewards
Computational checks (code execution, tool outputs) produce objective reward signals — no human feedback needed
LabBench2: verifiable tasks as RL training signal
Laurent et al., LabBench2, 2026; Narayanan et al., Aviary, 2024; 2025 NVIDIA Technical Blog
Behavior Changes
Training Curve
Narayanan et al., Aviary: training language agents on challenging scientific tasks, 2024
FutureHouse Agents
| Name | Environment | Key Tools |
|---|---|---|
| PaperQA | Literature Research | Search, Citation Traversal |
| ProteinCrow | Designing novel proteins | AlphaFold2, Molecular Dynamics |
| ChemCrow/Phoenix | Designing new molecules | Retrosynthesis, self-driving robotic lab |
| Data analysis crow | Generating discoveries from data | bioinformatics tools, code, file system |
Automating research of scientific literature
Language agents achieve superhuman synthesis of scientific knowledge
Michael D. Skarlinski, Sam Cox, Jon M. Laurent, James D. Braza, Michaela Hinks, Michael J. Hammerling, Manvitha Ponnapati, Samuel G. Rodriques, Andrew D. White arXiv:2409.13740, 2024
Measurement: LAB-Bench (2024)
Approximately what percentage of Drosophila with a H3.3K36R mutation finish developing and enclose?
- 80%
- 19%
- 50%
- 37%
- 6%
- 94%
Better at answering questions than PhD biology experts
Improving over time
Better than human written Wikipedia articles
Edison Literature (PaperQA3)
Added patents, clinical trials, opentargets, chinese-language preprints. Designed for scientists
- Provenance of claims
- All actions auditable
- Reproducible
PaperQA3 (2026)
Complete cycle of disease to target to drug
ROBIN: A Multi-Agent System for Automating Scientific Discovery Nature, 19 May 2026
Ali Essam Ghareeb*, Benjamin Chang*, Ludovico Mitchener, Angela Yiu, Caralyn J. Szostkiewicz, Jon M. Laurent, Muhammed T. Razzak, Andrew D. White†, Michaela M. Hinks‡, Samuel G. Rodriques
Robin proposes follow-up experiments
Robin optimizes the therapeutic candidates
State of Agents in 2026
Tool-State Paradox
$T_1=$ alphafold2
$T_2=$ $\beta$-sheet percentage
$T_2\rightarrow$ error, no pdb
$T_1,T_2\rightarrow$ success
$T_1,T_1,T_2\rightarrow$ need to treat multiple structures
Engineers must consider $O(2^{N})$ interactions between tools: will not scale past ~10 tools
Code-writing agents
LLMs manage state, you must only write good code
Challenges:
- Must create persistent sandboxed computers for agents
- Must solve security and authentication for agents
Kosmos Agent
A persistent always-on colleague that has access to 80% of public data, our chemistry and literature agents, and world models of biology
Google vs Agents
Agent
Create a dashboard of all trials, patents, major papers, and KOLs for γδ-T-cell programs. Email me every morning at 9am with updates. Send me a weekly podcast at 5pm on Thursdays with an overview.
Create a PowerPoint presentation giving an overview of changes in AEs unique to trials involving amylin analogs with GLP1s
Pull internal preclinical PK data on trevogrumab, propose 3 possible dosage schedules, and create a PDF report justifying them
Treat agents like an energetic and easily misguided colleague
Scientific Reasoning Models
Training a Scientific Reasoning Model for Chemistry
Siddharth M. Narayanan, James D. Braza, Ryan-Rhys Griffiths, Albert Bou, Geemi Wellawatte, Mayk Caldas Ramos, Ludovico Mitchener, Samuel G. Rodriques, Andrew D. White NeurIPS, 2025
Improving Models
| Pretraining | Large Data, Large Compute |
| Scaffolding | Domain knowledge |
| RL with verifiable rewards | Domain knowledge, small data, small compute |
Can we build scientific reasoning models?
chemistry reasoning model
Works with molecular structures, but reasons in English
Start from base LLM and teach it chemistry
What can a reasoning model do?
Q:Propose a 1-step synthesis path that uses only commercially available reagents
Q: Propose a modification to this molecule to increase its solubility by about 1 LogS unit without affecting its scaffold.
data
| Task | Subtasks | Examples | Verifier | Templates | Data source name |
|---|---|---|---|---|---|
| functional group | 1 | 74562 | code | 6 | ChEMBL |
| organism molecular formula | 1 | 74164 | molecule comparison | 10 | COCONUT |
| IUPAC name | 1 | 74994 | code | 10 | COCONUT |
| SMILES completion | 1 | 74990 | code | 10 | COCONUT |
| solubility edit | 3 | 115977 | ML model, code | 15 | ChEMBL |
| scent | 180 | 4240 | multiple choice | 8 | pyFUME |
| reaction prediction | 1 | 61205 | molecule comparison | 10 | ORD |
| retrosynthesis | 1 | 67252 | ML model, database | 8 | mcule |
| BBB permeability | 2 | 2064 | multiple choice | 8 | BBB |
| pKa | 4 | 336 | multiple choice | 8 | IUPAC |
| safety | 11 | 5687 | multiple choice | 8 | Pubchem |
| molecular formula | 1 | 18738 | code | 10 | COCONUT |
| ADME | 12 | 1030 | multiple choice | 8 | Fang ADME |
| LD50 | 2 | 342 | multiple choice | 8 | Pubchem |
| Human receptor binding | 150 | 1663 | multiple choice | 8 | EveBio |
| property-regression-solubility | 2 | 464 | multiple choice | 8 | AqSolDB |
| property-regression-photo | 1 | 23 | multiple choice | 8 | Photoswitches |
| Total | 374 | 577790 | 8 | 81* | 12 |
Training Stages
Results vs humans and frontier models
More data efficient
