The Age of AI Agents

How AI Agents Are Changing the Way We Do Science

CATH Lab Meeting • March 2026

The ChatGPT Era

🧑 "How do I parse a CIF file in Python?"
🤖 "You can use the Biopython library..."
⬇️ copy & paste into editor
🧑 "It's not working, I get this error..."
🤖 "Try changing line 12 to..."
⬇️ copy & paste again

You are the integration layer.

The Problem

  • You must sit in front of your computer
  • Context is lost between sessions
  • One task at a time
  • Copy-paste is the integration layer

What if AI could do the work,
not just answer questions?

Chat vs. Agent

💬 Chat

You drive, AI assists

You type → AI responds → You act

🤖 Agent

You delegate, AI executes

You assign → AI plans → AI acts → You review

From conversation to delegation

The Tool Landscape

⚡

Claude Code

  • Literature research
  • Data analysis
  • Presentation creation
  • Code review & debugging
🔧

Codex

  • Autonomous code writing
  • Test generation
  • Refactoring at scale
  • Multi-file changes

Live Demo: Claude Code

claude — ~/research
click to advance

Live Demo: Codex

codex — ~/protein-design
click to advance

Multi-Agent Collaboration

Example: Protein Design Pipeline

📚
Literature
Agent
Reads papers, extracts design principles
→
📥
Data
Agent
Downloads PDB structures, DMS datasets
→
🔍
Feature
Agent
Parses features, identifies hotspots
→
💻
Dev
Agent
Writes code, configures env, trains models
→
📝
Report
Agent
Generates figures, writes summary report

Each agent is specialized. They collaborate autonomously to complete the full pipeline.

The Key Shift

Before

You → Think → Code → Debug →
Test → Fix → Repeat
You are in the loop
→

After

You → Delegate
Agent → Plan → Execute → Verify
You are at the decision layer
OpenClaw

Your Personal AI Agent Hub

An agentic framework that connects various models,
e.g. Claude, Codex, GLM-5, MiniMax, etc.
Controllable via Telegram, Slack, WhatsApp, and more.

Architecture

📱 You Telegram 🐾 OpenClaw Agent Hub ⚡ Claude Research & Analysis 🔧 Codex Code Generation

click to animate data flow

The Freedom

I spend less time in front of my computer.

Anytime, anywhere — pull out your phone, send a task.

The agent works while you think.

OpenClaw Bot
Analyze LCC DMS data
Starting analysis... ✓
Complete. Report saved.

OpenClaw in Action

🐾
OpenClaw Bot
online
OpenClaw is typing...
click to advance

Mission Control

🧬
RBX1 De Novo Binder Design
RFdiffusion → AF3 → Protenix rescoring pipeline
📊
LCC DMS Data Analysis
Deep mutational scanning — mutation impact scoring
⚗️
Enzyme Specificity with CADD
Substrate docking + specificity profiling
🔬
Antibody Design
Generative model → structure prediction → optimization

RBX1 De Novo Binder Design

Target Selection
RBX1 RING domain
→
Binder Generation
RFdiffusion
→
Structure Prediction
AlphaFold 3
→
Rescoring
Protenix
→
Results
Top candidates

Files: design_rbx1_binder.py · rbx1_binder_report.html

← Back to Dashboard

LCC DMS Data Analysis

Raw DMS Data
→
Literature Research
→
Hot/Coldspot
Identification
→
DMS
Visualization
DMS Analysis ← Back to Dashboard

Enzyme Specificity with CADD

Enzyme Target
→
Substrate Docking
CADD
→
Specificity
Profiling
→
Design
Candidates

Orchestrated by OpenClaw's enzyme-specialist skill

← Back to Dashboard

Antibody Design

Antigen Input
Target name
→
Generative Model
Ab generation
→
Structure
Prediction
→
Affinity
Optimization

Agent receives an antigen name → produces candidate antibody sequences

← Back to Dashboard

Mission Control

🧬
RBX1 De Novo Binder Design
✓ 12 candidates generated, top 3 validated
📊
LCC DMS Data Analysis
✓ 3 significant mutations identified
⚗️
Enzyme Specificity with CADD
✓ 5 high-specificity variants found
🔬
Antibody Design
✓ 8 candidate antibodies, top 2 nanomolar

All handled by agents while you focus on science.

What This Means for You

  • You can build OpenClaw, but please start with Claude Code or Codex first (beginner-friendly)
  • Automate the tedious, focus on the creative
Thank You

Questions?

Weining Lin · zczlwl3@ucl.ac.uk

The future is agentic. 🚀