Session 01 · No-Code & Agentic AI for Life Sciences

No-Code Tools Setup
+ Foundations

Course overview · AI evolution · LLMs · Agentic stacks · Tool setup

Md. Jubayer Hossain · Founder & CEO
DeepBio Limited · DeepBio Academy

No Code & Agentic AI for Life Sciences · Session 01 · June 2026

Course Overview

What This Course Is

Hands-on program for life scientists who want to use AI agents — without writing production code.

  • No prior ML or coding required
  • Practical, reproducible research workflows
  • Tools used by researchers today
  • Privacy-first approach for sensitive data
  • Builds toward multi-step agentic pipelines

🎯 Who It's For

Biologists, clinicians, bioinformaticians, and data managers automating research tasks with AI agents.

📦 What You'll Build

Literature pipelines, data analysis workflows, scientific writing agents, custom skill-driven assistants.

🔒 Data Handling

Local-first where possible. Clear guidance on what leaves your machine and when.

Session 01

Today's Agenda

Part 1 — Concepts

  • AI evolution: ML → Agentic AI
  • Agentic AI landscape for life sciences
  • How LLMs work in practice
  • Agent-provider-model stack
  • Chat vs coding agents vs pipelines
  • Introduction to skills
  • Capabilities, limitations & failure modes
  • Data privacy: local vs external

Part 2 — Setup & Practice

  • VS Code / Claude Code / OpenCode setup
  • Install Scientific Agent Skills
  • Install Claude Scientific Writer Skills
  • Caveman setup for token management
  • First interaction exercise
  • Tool comparison exercise

⏱ Duration

~3 hours with 15-min break · 50% hands-on

Section 01

The AI Evolution

From statistical models to autonomous agents — knowing the arc helps you pick the right tool.

AI Evolution

ML → DL → GenAI → Agentic AI

📈
Machine Learning
Pattern recognition from labeled data. Rules learned, not written.
🧠
Deep Learning
Neural networks at scale. Image, speech, sequence tasks.
✨
Generative AI
Models that create: text, images, code, protein structures.
🤖
Agentic AI
Plans, uses tools, completes multi-step tasks autonomously.
"The shift from GenAI to Agentic AI is the shift from answering questions to completing tasks."

Why It Matters for Life Sciences

Agents can search literature, run analyses, write reports, and iterate — all from a single instruction.

Agentic AI Landscape

Agents in Life Sciences

📚

Literature & Discovery

PubMed pipelines, hypothesis generation, systematic review automation, citation management.

🧬

Omics & Analysis

RNA-seq workflows, genomic annotation, pathway analysis, structure prediction, variant interpretation.

📝

Scientific Writing

Manuscript drafting, abstract generation, methods rewriting, grant narrative assistance.

🔬

Lab & Protocols

Protocol optimization, reagent lookup, lab notebook automation, experimental design.

🏥

Clinical & Regulatory

Trial matching, adverse event summarization, regulatory document drafting, ICD coding.

📊

Data & Visualization

EDA automation, figure generation, statistical reporting, dashboard creation.

Section 02

Understanding LLMs

What large language models are, how they generate text, and what that means in practice.

LLMs

What LLMs Are & How They Work

LLMs are next-token predictors trained on massive text corpora. They learn statistical relationships between words, concepts, and reasoning patterns.

  • Trained on trillions of tokens of text
  • Generate responses token-by-token
  • No memory between sessions by default
  • Context window = working memory
  • Temperature controls randomness
  • System prompts shape behavior
# What the model actually does input: "The protein folds into a..." model: P(next_token | all_prior_tokens) output: "stable beta-sheet conformation"

Context Window

Everything the model can "see" at once — your prompt, conversation history, pasted documents. Measured in tokens (~¾ of a word).

Hallucination

LLMs generate plausible-sounding text, not verified facts. Always cross-check citations, gene names, drug interactions, and statistics.

For Science: Rule of Thumb

Use LLMs for reasoning, synthesis, writing. Use databases (PubMed, UniProt, PDB) for ground-truth facts.

Architecture

The Agent–Provider–Model Stack

Three distinct layers — swappable independently. Understanding them controls costs and privacy.

🤖
Agent / Interface
Tool you interact with. Sends prompts, receives responses, executes actions.
Claude Code, VS Code Copilot, OpenCode, Cursor
↕
🔌
API Provider
Routes requests to models. Handles auth, rate limits, billing.
Anthropic API, OpenRouter, OpenAI API, Azure AI
↕
🧠
Model
Neural network doing the work. Different sizes, costs, capabilities.
Claude Sonnet 4.6, GPT-4o, Gemini 2.5 Pro, Llama 3

Why This Matters

One agent, multiple models via different providers. OpenCode + OpenRouter routes to Claude, GPT, or Gemini from one interface.

Cost Control

Frontier models (Opus, GPT-4o) for complex reasoning. Fast cheap models (Haiku, GPT-4o-mini) for simple tasks. Same agent, different model.

Privacy Implication

Data travels to provider's servers on every API call. Local models (Ollama) keep everything on-machine.

System Types

Chat vs Coding Agents vs Agentic Pipelines

💬

Chat Assistants

Single-turn or conversational. You ask, it answers. No tool use or file access by default.

e.g. Claude.ai, ChatGPT, Gemini

Best for: Q&A, drafting, brainstorming

⌨️

Coding Agents

Read & edit files, run commands, use tools. Operate within a project context.

e.g. Claude Code, GitHub Copilot, Cursor

Best for: Analysis scripts, data workflows

⚙️

Multi-step Agentic

Agents calling agents. Plan → execute → evaluate → iterate. Minimal human-in-the-loop.

e.g. Claude Code + skills, custom pipelines

Best for: Full research workflows

This course focuses on coding agents and building toward multi-step agentic systems for life science research.

Skills

Introduction to Skills

A skill is a reusable instruction set that teaches an agent how to do a specific task well — like a detailed SOP loaded into the agent's context.

  • Plain-text files (.md) with structured instructions
  • Invoked with a slash command: /pubmed-search
  • Can call external APIs and databases
  • Composable — one skill can invoke another
  • Community-maintained and shareable
# Invoke a skill /pubmed-database "CRISPR off-target 2024" # Chain skills together /literature-review "mRNA vaccines" /scientific-writing methods section

🧪 Scientific Agent Skills

144 skills · 18 domains · 100+ databases. Bioinformatics, drug discovery, clinical research, lab automation, scientific writing. 27.9k ⭐ · 160k+ scientists.
K-Dense-AI/scientific-agent-skills

✍️ Scientific Writer Skills

Manuscript drafting, abstract writing, methods rewriting, figure legends, peer review responses. Included in Scientific Agent Skills.

🪨 Caveman Skills

Token management — avg 65% output reduction, 100% technical accuracy. Separate install.
juliusbrussee/caveman

Section 03

Strategic Frameworks

Bridging the gap between knowing the tools and using them safely, accurately, and affordably.

Critical Thinking

Capabilities, Limitations & Failure Modes

✓ Strong At

  • Synthesizing large text volumes
  • Structured writing and editing
  • Code generation & debugging
  • Explaining complex concepts
  • Brainstorming hypotheses
  • Following detailed instructions

⚠ Unreliable For

  • Specific citations (hallucinated DOIs)
  • Real-time data (cutoff: Aug 2025)
  • Precise numerical calculations
  • Novel experimental design alone
  • Regulatory / legal conclusions
  • Rare disease edge cases

✗ Known Failure Modes

  • Confident hallucination of facts
  • Sycophancy (agreeing too readily)
  • Context window overflow / drift
  • Instruction degradation over long sessions
  • Prompt injection via pasted text
  • Inconsistency across sessions
"Treat LLM output as a first draft from a well-read but sometimes overconfident assistant. Always verify against primary sources."
Trust but Verify

Verification & Ground Truth

Agents execute code and synthesis, but the scientist provides the Audit. Never accept an agentic result without a validation step.

The "Agent Auditing" Workflow

  • Show Your Work: Require the agent to print intermediate dataframes or logs.
  • Source Triangulation: Cross-check agent summaries against primary PDFs or PubMed DOIs.
  • Reasoning Trace: Ask: "Explain the logic used to prioritize these 5 genes."
  • Null Hypothesis Test: Ask the agent to find evidence against its own conclusion.

🔍 The "Code Review" Rule

When an agent writes a script (e.g., for RNA-seq), read the generated code before execution. Check for hardcoded assumptions or incorrect library versions.

🚫 The Hallucination Trap

Agents are most likely to hallucinate when forced to provide a specific fact they don't have. Use grounded search tools (PubMed/UniProt) to fix this.

Privacy

Data Privacy: Local vs External

🟢 Stays Local

  • Files on filesystem (not pasted in)
  • Local model inference (Ollama, LM Studio)
  • Conversation history in CLI (local DB)
  • Code run in terminal / sandbox

🟡 Sent Externally

  • Any text in your prompt / message
  • Files you paste or attach to context
  • Tool call inputs & outputs (via API)
  • Web fetch results within agent context

🏥 Clinical / Patient Data

Never paste identifiable patient data into any cloud-API-backed agent. Use de-identified summaries or local models only.

📋 Check Provider Policies

Anthropic API & Claude Code: data not used for training by default. Verify with your institution's data governance team.

Prompt Engineering

Framing Scientific Queries

Moving from "Chat" to "Directive." Better framing leads to higher-fidelity scientific reasoning.

The C-T-C Framework

1. Context (The Persona)

"Act as a senior computational biologist specialized in single-cell transcriptomics."

2. Task (The Directive)

"Evaluate the off-target profile of this gRNA sequence against the human GRCh38 genome."

3. Constraints (The Guardrails)

"Prioritize results with < 3 mismatches. Format as a table with Ensembl IDs and citations."

"The quality of the output is determined by the clarity of the constraints, not just the complexity of the task."

💡 Pro Tip: Multi-Shot

Give the agent 1-2 examples of the output you expect (e.g., a specific table format). This reduces trial-and-error costs.

Financial Literacy

Budgeting Your AI Assistant

Professional agents use Pay-Per-Token billing. Inefficient use can lead to "Bill Shock."

Cost-Saving Strategies

  • Caveman Mode: Use /caveman full to save ~65% on output tokens.
  • Context Management: Use /forget or start new sessions to clear "heavy" history.
  • Billing Alerts: Set a $10 hard limit in the Anthropic/Google console.
  • Model Tiering: Use "Flash" or "Haiku" models for simple tasks; save "Pro" or "Opus" for complex reasoning.

💰 Typical Costs

A full hour of agentic research usually costs $0.50 – $2.00. Writing a single paper draft may cost $5.00 – $15.00 depending on model tier.

🛡️ Safety First

Never hardcode API keys. Use Environment Variables (.env) and ensure they are in your .gitignore.

Section 04

The AI Toolset Landscape

A deep dive into the agents, environments, and scientific ecosystems used in this course.

Agentic CLI

Claude Code

Anthropic's "coding partner" that reasons across codebases, runs terminal commands, and manages complex multi-file tasks.

Core Workflow (Reason & Act)

  • Multi-File Reasoning: Maps directory structures and identifies cross-module dependencies.
  • Tool Execution: Natively runs git, npm, pytest, and custom bash scripts.
  • Iterative Debugging: Edits code → runs tests → reads logs → fixes until green.
  • Checkpointing: Automatic state saves with /rewind to undo complex refactors.

🛡️ Model Context Protocol (MCP)

Connects Claude to GitHub, Linear, Slack, and production databases to bring real-world context into the agent session.

⚡ Parallel Sessions

Supports git worktrees to handle a refactor, a feature, and a bug fix simultaneously in isolated environments.

🎯 Recommended For

Developing bioinformatics pipelines, refactoring analysis scripts, and complex data migrations.

Agentic CLI

Gemini CLI

An open-source agent built on Gemini 2.5 Pro specialized in grounded research and surgical code edits.

Agent Life Cycle

  • Research: Uses google_web_search and web_fetch to consult latest documentation.
  • Plan Mode: Dedicated read-only mode (/plan) for safe architectural design before writing.
  • Act (Surgical): High-precision replace tool for minimal diffs and efficient token usage.
  • Validate: Automated check-ins against project-specific GEMINI.md mandates.

🏗️ Hierarchical Memory

Uses GEMINI.md at global, project, and folder levels to maintain persistent style and architecture rules.

🌐 Multimodal Support

Can "see" images and screenshots (Ctrl+V) to debug frontends or interpret scientific charts.

🎯 Recommended For

Grounding research in latest papers, medical imaging analysis, and large-context codebase mapping.

Scientific Multi-Agent

Co-Scientist (DeepMind)

A multi-agent system designed to propose, debate, and refine novel scientific hypotheses.

The "Tournament of Ideas" Workflow

  • Generation: Supervisor agent coordinates specialized agents to propose focus areas.
  • Debate: Reflection agent acts as virtual peer reviewer, ranking ideas in an Elo-based tournament.
  • Evolution: Top-ranked hypotheses are iteratively refined and combined for robustness.
  • Synthesis: Meta-review agent generates a comprehensive research overview.

🔬 Real-World Success

Replicated 10 years of antimicrobial resistance research in 48 hours.

🔗 Tool Chaining

Natively invokes AlphaFold for structure prediction and UniProt for sequence analysis.

🎯 Recommended For

Hypothesis generation, antimicrobial discovery, and drug repurposing research.

Biomedical Agent

Biomni (Stanford)

A "virtual AI biologist" capable of autonomously executing tasks across 25 biomedical subfields.

Architecture: E1 (Env) + A1 (Agent)

  • Unified Action Space: Integrated with 150 tools, 105 software packages, and 59 databases.
  • CodeAct: Uses Python code as the universal interface for tool invocation and statistical analysis.
  • Judge/Actor Loop: A "Judge" LLM verifies every step to prevent confabulation (hallucination).
  • Performance: Achieved 74.4% on LAB-Bench, rivaling human expert performance.

🧪 Know-How Library

Retrieves 10-step experimental protocols (e.g., molecular cloning) rated at senior-scientist level.

🧬 Clinical Reasoning

Specialized in GWAS interpretation, causal gene prioritization, and rare disease diagnosis.

🎯 Recommended For

Experimental protocol design, clinical genomics, and causal inference in biology.

Scientific Ecosystem

K-Dense.ai Ecosystem

The foundational framework powering the Scientific Agent Skills used throughout this course.

Core Components

  • Skills Library: 144+ composable skills covering 18 domains (Omics, Chem, Clinical).
  • Expert Thinking: 500+ agent profiles encoding the reasoning patterns of senior scientists.
  • Science-Superpowers: Methodology skills that enforce pre-registration and hypothesis validation.
  • BixBench Leader: Outperforms raw models by 50%+ in bioinformatics tasks via hierarchical loops.

🏗️ Open Agent Skills

Native compatibility with Gemini CLI, Claude Code, Cursor, and any OAS-compliant agent.

📂 Mimeo

A specialized tool to "mimeograph" human expertise into structured agent-ready SOPs.

🎯 Recommended For

Building custom multi-agent pipelines and standardizing lab-specific agentic workflows.

Section 05

Environment & Setup

Final technical preparation and installation of the course toolkit.

Environment

The "Health Check" Prerequisites

Before we begin the setup, ensure your local environment is ready for agentic tools.

📦 System Packages

  • ✓ Node.js 18+: Required for Claude Code.
  • ✓ Python 3.12+: Required for Scientific Agent Skills.
  • ✓ Git: Installed and user.email configured.
  • ✓ Terminal: VS Code Terminal, iTerm2, or PowerShell 7.

🔑 API & Access

  • ✓ Anthropic API Key: Funded with min. $5.
  • ✓ Google Cloud Key: For Gemini CLI features.
  • ✓ GitHub Account: For Copilot and skill discovery.
  • ✓ Stable Internet: High-bandwidth for model streaming.

Ready to Install? Proceed to Tool Options.

Tool Setup

Three Setup Paths

🐙

VS Code + GitHub Copilot

Already in VS Code ecosystem. Subscription-based. GPT-4o / Claude Sonnet backed.

Requires: VS Code, GitHub account, Copilot (~$10/mo)

⚡

Claude Code Recommended

Deep research workflows, rich skills ecosystem, powerful agentic mode. Terminal + IDE.

Requires: Node.js, Anthropic API key. Usage-based billing.

🌐

OpenCode + OpenRouter

Model flexibility and cost optimization. One interface → Claude, GPT, Gemini, Llama.

Requires: OpenRouter API key. Pay-per-token across models.

Recommended for this course: Claude Code — richest skills ecosystem, tightest integration with scientific agent skills we'll install. Works in terminal and VS Code.
Setup

Install Scientific Agent Skills

144 skills · 18 domains · 100+ databases · 27.9k ⭐ · used by 160,000+ scientists · K-Dense-AI/scientific-agent-skills

Install (Python 3.13+, uv required)

# Install uv first (all platforms) curl -LsSf https://astral.sh/uv/install.sh | sh # NPX — recommended, works everywhere npx skills add K-Dense-AI/scientific-agent-skills # GitHub CLI alternative gh skill install K-Dense-AI/scientific-agent-skills

Works with Claude Code, Cursor, Codex, Gemini CLI, and any Agent Skills-compatible tool. Security scanned via Cisco AI Defense.

# Example: drug discovery workflow "Query ChEMBL for EGFR inhibitors, analyze with RDKit, screen with DiffDock, check COSMIC, create report" # Example: single-cell analysis "Load 10X data with Scanpy, integrate Cellxgene Census, run DESeq2"

Skill Domains

🧬 Bioinformatics & Genomics (23)

Scanpy, BioPython, pysam, scVelo, PyDESeq2, RNA-seq, single-cell analysis

💊 Cheminformatics & Drug Discovery (10)

RDKit, Datamol, DiffDock, DeepChem, molecular dynamics, OpenMM

🏥 Clinical & Precision Medicine (8)

ClinicalTrials.gov, ClinVar, DepMap, clinical decision support, COSMIC

✍️ Scientific Communication (27)

PubMed, literature review, scientific writing, peer review, presentations

🗄️ Databases (100+)

PubChem, ChEMBL, UniProt, KEGG, STRING, Ensembl, FDA, ClinVar, FRED, USPTO

Setup

Claude Scientific Writer Skills

Two-stage pipeline: outline → full prose. Publication-quality text with citations. K-Dense-AI/claude-scientific-writer

Install

# NPX — recommended, works everywhere npx skills add K-Dense-AI/claude-scientific-writer # GitHub CLI alternative gh skill install K-Dense-AI/claude-scientific-writer

Usage

# Invoke the writing skill /scientific-writing # With a direct task /scientific-writing "Write the methods section for RNA-seq differential expression analysis using DESeq2"

Skill uses research-lookup internally. Outputs IMRAD structure, APA/AMA/Vancouver citations.

Writing Skill Capabilities

📄 Full Manuscripts

Introduction → Methods → Results → Discussion. Section-by-section or full draft.

🔬 Methods Sections

Reproducible methods from protocols. Reagent lists, statistical parameters, software versions.

📋 Reporting Guidelines

CONSORT / STROBE / PRISMA compliance built-in. Checklist-aware generation for clinical studies.

💬 Peer Review Responses

Point-by-point reviewer response letters with rebuttal arguments and manuscript edits.

Token Management

Caveman: 65% Output Reduction

Caveman is a Claude Code skill that compresses agent output ~65% while keeping 100% technical accuracy. github.com/juliusbrussee/caveman

Install (requires Node ≥18)

# macOS / Linux / WSL curl -fsSL https://raw.githubusercontent.com/ JuliusBrussee/caveman/main/install.sh | bash # Windows PowerShell 5.1+ irm https://raw.githubusercontent.com/ JuliusBrussee/caveman/main/install.ps1 | iex

~30 sec install. Re-runnable safely. Auto-activates for Claude Code, Codex, Gemini.

Before vs After

# Normal (1,180 tokens for React bug) "The reason your React component is re-rendering is likely because you're creating a new object reference each render." # Caveman (159 tokens — 87% saved) "New obj ref each render. Inline object prop = new ref = re-render. Wrap useMemo."

⛏ Slash Commands

/caveman [lite|full|ultra|wenyan] — set compression level
/caveman-stats [--share] — session tokens saved + USD cost
/caveman-commit — conventional commits ≤50 chars
/caveman-review — one-line PR feedback with emoji
/caveman-compress <file> — compress memory files ~46%

Compression Levels

lite — drop filler text only
full — default (fragments, minimal conjunctions)
ultra — telegraphic, maximum brevity
wenyan — classical Chinese encoding (most compact)

Real Benchmarks

React bug: 1,180 → 159 tokens (87% saved)
Auth fix: 704 → 121 tokens (83% saved)
PostgreSQL setup: 2,347 → 380 tokens (84% saved)
Memory files avg: 46% reduction

Hands-On

First Interaction & Tool Comparison

Exercise 1 — First Interaction

  1. Open your chosen agent (Claude Code / Copilot / OpenCode)
  2. Run: "Summarize what CRISPR-Cas9 is in 3 sentences for a non-specialist."
  3. Then: "Add 2 recent applications in oncology."
  4. Observe: Did it remember context? Hallucinate any studies?
  5. Verify one claim against PubMed

Exercise 2 — Tool Comparison

  1. Run the same CRISPR prompt in two different tools
  2. Compare: accuracy, speed, style, token usage
  3. Try /pubmed-database: "Find 3 reviews on CRISPR off-target effects"
  4. Compare skill result vs plain chat result

✍️ Group Discussion

Which tool felt most useful and why? No right answer — context matters.

Summary

Session 01 Key Takeaways

  • AI evolved: ML → DL → GenAI → Agentic AI
  • Agents automate research end-to-end in life sciences
  • LLMs predict tokens — they don't "know" facts
  • Agent ≠ Provider ≠ Model — layers are separable
  • Chat, coding agents, pipelines: different use cases
  • Skills = reusable instruction sets for specific tasks
  • Hallucination is real — always verify critical claims
  • Data privacy depends on where inference happens
  • Caveman mode reduces token cost on long sessions

📅 Next Session

Session 02: Literature Search & Review Automation — PubMed agent workflows, systematic review pipelines, citation management.

📚 Before Next Session

Complete tool setup. Run first interaction exercise. Explore /get-available-resources and note 3 skills relevant to your research.

❓ Questions & Support

Office hours: Thursdays 16:00–17:00 BST.
Email: work.jubayer@gmail.com

No Code & Agentic AI for Life Sciences

Thank You

Get in Touch

Md. Jubayer Hossain

bio.link/hossainlab

© 2026 Md. Jubayer Hossain · No Code & Agentic AI for Life Sciences — Session 01