This is a submission for the Google I/O Writing Challenge
The Moment That Stopped Me Mid-Scroll
I was half-watching the Google I/O 2026 keynote during a break between lectures when something Sundar Pichai said made me put my coffee down. It was during the opening segment — Google's vision for AI assisting educators and learners. Not replacing them. Assisting them.
That hit differently for me because I'm not a developer by training. I'm a Public Health Biotechnology master's student, drowning in epidemiology formulas, statistical tests I can never remember the assumptions for, and primer design rules that change depending on whether I'm doing standard PCR or qPCR.
Then I switched to the Developer Keynote, and Josh Woodward started talking about building agents — not just chatbots, but agentic systems that reason, use tools, and collaborate. The session "Build agents with Gemini API" talked about going from prompt to production agent with the Gemini Interactions API. The "Future of Software Development" session described how "coding is evolving from writing boilerplate to curating outcomes."
And I thought: what if I'm not just a consumer of AI tools — what if I build the tools that help me learn?
So I did. And Google's own architecture validated every decision I made.
What Google Announced (and Why a Biotech Student Cared)
Let me ground this in the specific I/O 2026 announcements that matter here:
1. The Agentic Paradigm Shift
Google I/O 2026 made one thing abundantly clear: we've moved from "prompt → response" to "agent → autonomous reasoning → action." The Gemini API now supports managed agents — a single API call provisions a Linux sandbox where the agent reasons, executes code, manages files, and browses the web autonomously1.
But the announcement that made me sit up was Google's Agent Development Kit (ADK) 2.0 going GA with graph workflows and collaborative agents2. ADK introduced a concept called Skills — "self-contained units of functionality that an agent can use to perform a specific task"3. Skills follow a three-level progressive loading architecture:
- L1 (Metadata): Skill name and description for discovery
- L2 (Instructions): Primary procedure, loaded when triggered
- L3 (Resources): Reference materials, loaded on demand
Sound familiar? It should. Because I independently built the exact same architecture for my coursework — and I was doing it before ADK 2.0 was announced.
2. AI for Learners, Not Just Developers
The Google Keynote's framing of AI as a tool that "enhances creativity, extends knowledge, and advances science"4 wasn't just corporate messaging to me. It was a description of my Tuesday afternoon. I use AI to understand disease mechanisms, check my statistical reasoning, and generate structured slides for seminars. The difference is: I got tired of explaining my field to the AI every. single. time.
3. From Vibe Coding to Vibe Learning
The "Build next-gen AI experiences with Google AI Studio and Google Antigravity" session introduced Google's agent-first IDE where you "speed-run ideas in AI Studio, export code, and deploy agents to handle heavy lifting like planning architecture [and] writing multi-file features."5
I realised: what if the "heavy lifting" isn't code architecture — what if it's the cognitive architecture of a master's programme?
What I Actually Built
I built pub-health-biotech-skills — a modular repository of reusable AI skills, expert agents, and one-shot prompts covering my entire MSc curriculum. It follows the same L1/L2/L3 progressive loading pattern that Google's ADK now formalises as a specification6.
Here's the architecture:
pub-health-biotech-skills/
├── skills/ # Domain-specific workflows (progressive loading)
│ ├── epi-study-design/
│ │ ├── SKILL.md # L2: Procedure + triggers
│ │ └── references/ # L3: Loaded on demand
│ │ ├── study-types.md
│ │ └── bias-checklist.md
│ ├── biostat-test-selector/
│ │ ├── SKILL.md
│ │ └── references/
│ │ ├── decision-tree.md
│ │ └── assumptions.md
│ ├── molgen-primer-design/
│ ├── genetics-pedigree/
│ └── ... (13 skills total)
├── agents/ # Expert persona agents
│ ├── epidemiologist.agent.md
│ ├── biostatistician.agent.md
│ ├── geneticist.agent.md
│ └── molecular-biologist.agent.md
├── prompts/ # One-shot task templates
└── docs/ # Cross-LLM compatibility
Three Layers, Three Purposes
Skills are workflows. My biostat-test-selector skill doesn't just say "help with stats." It contains a step-by-step decision tree: ask the outcome type → check number of groups → determine independence → verify assumptions → recommend a test → provide the R code → give the APA reporting template. The reference files hold the actual formulas and assumption checklists — loaded only when needed.
Agents are expert personas with hard constraints. My @biostatistician refuses to use stepwise selection for causal models, won't interpret p > 0.05 as "no effect," and always reports effect sizes alongside p-values. My @epidemiologist won't let you calculate a relative risk from a case-control study. These aren't stylistic preferences — they prevent the kinds of errors that cost you marks on an exam.
Prompts are one-shot commands. Type /critique-study, paste the paper, get a structured evaluation against CARE or STROBE guidelines with scoring.
The Google I/O Connection: Why This Architecture Works
Here's what Google's announcements validated:
Progressive Loading Solves the Context Problem
ADK's documentation states that skills help "optimize the context window by only loading instructions when they are needed"3. I learned this the hard way. My first attempt loaded everything — all formulas, all decision trees, all guidelines — into every conversation. The AI's responses got worse. It was juggling Hardy-Weinberg equilibrium math while I was asking about study design bias.
Splitting into skill-specific references with on-demand loading fixed response quality immediately. Google formalised this as L1/L2/L3. I call it "not overwhelming the AI with my entire curriculum."
Multi-Agent Workflows Are the Future
ADK 2.0's collaborative agent workflows2 mirror what I'm building toward: chain the biostatistician and epidemiologist together to draft a complete methods section. Design → sample size calculation → analysis plan → reporting template, all in one flow. Google's architecture proves this isn't a hack — it's the direction the industry is heading.
Skills Are Reusable Across Models
The "What's New in Google AI" session showcased building with the full Gemini model family. My skills are model-agnostic markdown files. They work in VS Code with GitHub Copilot, in Claude Projects, as ChatGPT custom instructions, or as system prompts for Gemini. The knowledge doesn't care what model reads it — just like ADK supports Gemini, Gemma, Claude, Ollama, and others7.
A Real Workflow: Exam Revision at 11pm
It's the night before my biostatistics exam. I need to revise statistical test selection.
Without skills:
"You are a biostatistician. I have a continuous outcome, 3 independent groups, small sample size, probably not normally distributed. Which test should I use? Walk me through the decision process, check assumptions, give me R code, and format it in APA style. Also show the non-parametric alternative in case assumptions are violated. And don't forget to mention post-hoc tests..."
With skills:
/biostat-test-selector"Continuous outcome, 3 groups, n=15 each, non-normal"
The AI already knows the decision tree. It walks through it step by step. Recommends Kruskal-Wallis. Gives me the R syntax. Shows the post-hoc pairwise comparison approach. Provides an APA reporting template. 30 seconds, done.
That's not a time saving. That's a cognitive saving. At 11pm before an exam, the last thing I need is to spend mental energy crafting the perfect prompt.
My Honest Critique: What Google Got Right and What's Missing
What They Got Right
The Skills specification is brilliant. The three-level loading (metadata → instructions → resources) with the SKILL.md + references/ structure is exactly what domain-heavy applications need. It's not over-engineered — a SKILL.md file is just markdown with YAML frontmatter. Anyone can write one.
ADK is model-agnostic by design. Supporting Gemini, Claude, Ollama, and LiteLLM means skills aren't locked into Google's ecosystem. This matters for students who can't always pay for API access.
The emphasis on "curating outcomes" over "writing code" in the developer keynote resonates. I'm not an engineer. I'm a public health student who happens to know enough markdown to build an agent system. The barrier to entry is low, and that's intentional.
What's Missing (for learners like me)
No education-specific examples. Every agent demo at I/O was about travel booking, logistics, or enterprise. Where are the agents that help students learn? The architecture supports it perfectly — Google just hasn't shown it.
Skills discovery is still primitive. ADK Skills are triggered by description matching. For a student with 13 skills across 4 domains, the AI sometimes picks the wrong one. I want curriculum-aware routing: "I'm studying for my Epi Methods exam" → automatically surface only epidemiology skills.
No built-in validation against ground truth. When my @biostatistician agent recommends a test, I trust it because I wrote the decision tree from my textbook. But there's no mechanism to verify outputs against known-correct answers. For educational use, this matters more than for enterprise use.
Five Lessons for Anyone Building Learning-Focused Agents
Constraints produce better output than instructions. Telling my agent "do NOT interpret p > 0.05 as no effect" prevents more errors than "always interpret carefully." Google's managed agents use similar constraint patterns in their safety documentation1.
Structure beats length. A concise decision tree outperforms a 2,000-word system prompt. The AI follows procedures better than prose. ADK's skill structure enforces this with the L2/L3 separation.
Domain knowledge is the moat. The SKILL.md is routing logic. The reference files — the formula sheets, the bias checklists, the assumption verification steps — that's where the real value lives. Anyone can build an agent. Few people will curate domain-specific knowledge with enough rigour to be trustworthy.
Agent personas need opinions. A generic "statistics helper" gives wishy-washy answers. A biostatistician who insists on confidence intervals and refuses stepwise regression gives answers you can actually submit. Give your agents convictions.
Build for the model you'll use tomorrow, not today. My skills are markdown files with no vendor lock-in. When Gemini 3.5 Flash drops in price or a new open-source model appears, I just point it at the same skill files. Google's multi-model support in ADK confirms this is the right call7.
What's Next: The Roadmap
This project started as a productivity hack. After watching I/O 2026, I see it as something bigger.
Near-term: More reporting guideline skills (STROBE for observational studies, CONSORT for RCTs, PRISMA for systematic reviews). An exam revision mode that generates practice questions from lecture notes and marks my answers.
Medium-term: Multi-agent workflows using the patterns ADK 2.0 formalised — chain the epidemiologist and biostatistician for complete methods section drafting. Integration with PubMed for evidence-grounded responses. Interactive R notebooks that execute analyses, not just suggest code.
Long-term: Open this to other health science students. Immunology skills. Pharmacology skills. Nursing research methods. A community-contributed skill library where every student's domain expertise benefits everyone else — the same collaborative model Google is pushing with their open-source ADK repositories.
The Bigger Picture
Google I/O 2026's message was clear: we're moving from AI as a chatbot to AI as an autonomous collaborator. The tools are here — Gemini Interactions API, managed agents, ADK Skills, agent-first IDEs.
But the most underrated insight from the entire event wasn't technical. It was philosophical. In the keynote intro, when they talked about AI helping educators and learners, they weren't announcing a product. They were acknowledging a use case that developers often overlook: the people who need AI most aren't building software — they're building knowledge.
I'm a public health student. My "production environment" is an exam hall. My "deployment target" is a research proposal. My "users" are my supervisors who will read my thesis. And the agents I built help me meet those targets with more rigour, more consistency, and more confidence than prompting from scratch ever could.
Google gave us the architecture. The question is: what will you build with it that isn't another to-do app?
References
Disclosure: This article was written with assistance from GitHub Copilot (Claude Opus 4.6) for drafting and structural editing. All domain knowledge, project architecture, and opinions are my own. The project described is real and available on GitHub.
If you're a student building something similar for your field — or if you're a developer who's never thought about AI for education — I'd love to hear your take in the comments.
-
Google. "Agents Overview — Gemini API." Google AI for Developers, 19 May 2026. https://ai.google.dev/gemini-api/docs/agents ↩
-
Google. "ADK Python 2.0 GA — Graph Workflows and Collaborative Agents." Agent Development Kit Documentation, 2026. https://adk.dev/2.0/ ↩
-
Google. "Skills for ADK Agents." Agent Development Kit Documentation, 2026. https://adk.dev/skills/ ↩
-
Google. "Google I/O 2026 Writing Challenge — Sponsored by Google AI." DEV Community, May 2026. https://dev.to/challenges/google-io-writing-2026-05-19 ↩
-
Google. "Build next-gen AI experiences with Google AI Studio and Google Antigravity." Google I/O 2026 Session, May 2026. https://io.google/2026/explore/pa-keynote-2 ↩
-
Agent Skills Specification. https://agentskills.io/specification ↩
-
Google. "Agents — Models for Agents." Agent Development Kit Documentation, 2026. https://adk.dev/agents/models/ ↩
Top comments (0)