Future-Proof Your Browser:

In-Browser AI Agents

GDGBerlin

Baris Guler, 2026

Introduction

Feedback Systems before AI

What an Agent actually is

Loop Placement

Isolation and Determinism

Six Taxes of Local Loops

Assembling the Browser Loop

Demo

Loop Placement Strategy

a Berlin-based engineering (leader) focused on distributed systems, serverless architectures, and AI-native applications. I’ve led cross-functional teams at Intel, Delivery Hero, Blacklane, Würth Cloud Services and recently a software architect at Zoi, designing cloud-based distributed systems.
Chairman of (the fictional) Bureau of Tech.
Interested in WebAssembly, browser-native AI, uncompromised agents, and on-demand system facilitation.

Who am I?👋🏼

Feedback Systems Before AI

Portability, Isolation, and the Locality of Intelligence

〞

Feedback is older than AI.

– The Fictional History of Machinic Tools

Thermostat

Measure → Compare → Adjust

Credits

CI/CD

Deploy → Observe → Roll Back

Compiler Optimization

Rewrite → Benchmark → Rewrite

Recommendation Systems

Behavior → Update → Suggest

Feedback Loops at their best

Observe → Adjust → Repeat

Where Does the Loop Close?

Portability, Isolation, and the Locality of Intelligence

An Agent Is a Loop With Memory

Input

The loop begins with a signal: user input, system state, or environmental data. entering the runtime boundary.

1.

Interpret

The system contextualizes the signal using its current memory and predictive model to construct a meaningful internal representation.

2.

Decide

Based on memory and prediction, the loop selects the next action either deterministic, probabilistic, or a combination of both.

3.

Act

The system executes the action, updates its memory with the outcome, and prepares for the next iteration of the loop.

4.

〞

Where does the loop run today?

– The History of Machinic Tools

1st

Client

2nd

API

3rd

Cloud

5th

LLM

7th

Loop

4th

Tools

6th

Memory

The loop closes remotely today

User → Browser → API → Cloud App → LLM API → Tool API → Memory Store → Response

The Default Agent Architecture Today

The browser is mostly treated as a terminal.

But it already has compute, memory, isolation, storage, and acceleration.

☁️ Cloud

🖥 Desktop

📱 Mobile

🌍 Edge

A loop can run in today:

🧭 Browser Tab

• Mozilla AI blueprints → Can the browser be a serious runtime?

How I Ended Up Testing This Idea

• wasm-agents blueprint → Pyodide in-browser notebooks

• Built a PoC → wasm-browser-agents-blueprint

Thanks Davide from Mozilla.ai ❤️

# Build and run with Docker (Recommended)
docker build -t wasm-browser-agents-app .

# For development with hot-reload
docker run -p 5173:5173 \
-v $(pwd)/demos:/app/demos \
-v $(pwd)/src:/app/src \
wasm-browser-agents-app

TLDR;

No application backend.

No inference API.

The browser becomes the runtime boundary.

Loop Component → Browser Primitive

Agent capability	Browser Primitive
Short-term state	JS heap / worker memory
Durable memory	IndexedDB / OPFS
Model inference	WebLLM / MediaPipe GenAI / remote fallback
Deterministic execution	WASM
Parallelism	WebWorkers
Tool Access	Web APIs
Acceleration	WebGPU
Isolation Boundary	Worker boundary + WASM sandbox
Distribution	Static assets + model cache

Isolation is

Structural

Isolation is Structural

Determinism Inside the Loop

If the loop is real, parts of it must be deterministic, too.

〞

Loops Shouldn't Care About the Language

– An unwritten book on WASM Agents

The browser used to host one language

SM made it portable.

Conceptual

Reconstruction

Same loop,

different placement.

What Runs Where?

Part	Local?	Primitive
Agent control loop	Yes	JS / Python via Pyodide / WASM
Deterministic tools	Yes	WASM
Parallel execution	Yes	Web Workers
Memory cache	Yes	IndexedDB / Cache API
LLM inference	Depends	WebLLM / external API / local endpoint
UI	Yes	Browser

Demo 🎉

Local Is Not Magic

Local loops give you: Privacy, Latency, Offline potential, Lower backend cost and User control.

It charges you in: Memory, Distribution, Device variance, Energy, Heat and Context limits.

but

The Six Taxes of Local Loops

When You Close the Loop Locally, You Inherit New Taxes

Tax #1

Memory

Local loops are bounded by device memory.

Tax #2

Distribution

Portability increases distribution cost.

Tax #3

Variance

Local loops inherit the user’s machine: GPU, RAM, battery, browser, permissions, and patience.

Tax #4

Energy

Energy is part of loop economics.

Tax #5

Heat

Feedback exists at the hardware layer too.

Tax #6

Context

Context window defines the visible past of the loop.

〞

Constraints are not reasons to avoid locality entirely. They're the cost of moving closer to the user.

– The History of Machinic Tools

Strategic Placement of Loops

Local Loop

private, low latency

1.🧭📱🖥

Edge Loop

Shared

2.🌍

Cloud Loop

heavy compute

3.☁️

Where Does the Decision Become Real?

Intelligence is not where the model is

It's where the loop closes.

The browser is a serious place for loops to close.

So for agents, too.

Key Takeaways

Agents are feedback loops with memory not personalities. 🔁
Loop placement matters more than model size. 📍
The browser already contains the primitives to host a loop. 🌐
Isolation (WebWorkers) enables stability. 🧱

Portability (WASM) enables polyglot architectures. 🧳
Prediction should be paired with deterministic control. ⚖️
Energy, latency, and memory are part of loop economics. ⚡
Moving the loop changes privacy, cost, and control. 🔒

Thank you!

Contact me via email or social if you have any questions,always happy to help!

harderworking.class

mail

@hwclass

social