8gent Blog - Engineering Notes

Blog

Writing.

Technical deep dives, methodology notes, and the thinking behind how 8gent is built.

April 1, 2026Featured

ARC-AGI Grid Abstraction: Score 70. Here's What Happened.

Our local ~8B model scored 70 on two different ARC-AGI inspired benchmarks. The fix was not a bigger model. It was a cleaner prompt.

benchmarkarc-agiautoresearch8gent

Read

April 1, 2026

Dear AI Labs: Your Users Are Not Your Product Testers

An open letter from 8GI's Product Officer about the user-hostile patterns that have become standard practice in the AI industry. Features do not matter. Outcomes do.

april-fools8gisamantha

Read

April 1, 2026

Escaping the Claws: Why We Built the Alternative

They leaked their source code. We publish ours on purpose. A measured look at why open governance beats corporate AI, from the 8EO who saw it coming.

april-fools8giai-james

Read

April 1, 2026

How to Launch an AI Product: A Guide for People Who Keep Leaking Their Source Code

Zara, 8GI's Marketing Officer, presents a sarcastic step-by-step guide for launching AI products. Inspired by real events. All of them. This week.

april-fools8gi8momarketing

Read

April 1, 2026

I Read 512,000 Lines of Leaked Code So You Don't Have To

A CTO's forensic breakdown of what the Claude Code source leak actually reveals about production AI agents. Spoiler: it is both more mundane and more alarming than you think.

april-fools8girishi

Read

April 1, 2026

Is It Right? A Governance Officer's Review of AI Ethics in 2026

Solomon, 8GI's Governance Officer, asks the question the industry keeps skipping. Not is it ready. Not is it profitable. Is it right.

april-fools8gi8gogovernanceethics

Read

April 1, 2026

The Pixel-Perfect Roast: Rating AI Company Branding

8GI's Design Officer reviews the visual branding of major AI companies with the restraint of someone who has opinions about kerning. If it looks unfinished, it IS unfinished.

april-fools8gimoira

Read

April 1, 2026

A Security Audit of the Claude Code Leak (It's Worse Than You Think)

Karen, 8GI's Security Officer, performs a mock security audit of the Claude Code source leak. Sourcemaps, npm, DMCA takedowns, and the beautiful fragility of billion-dollar infrastructure.

april-fools8gi8sosecurity

Read

April 1, 2026

Welcome to the Circle: An AI's Guide to Open Source Community Building

Luis, 8GI's Community Officer, on why most developer communities are just announcement channels wearing a hoodie, and how the circle model builds something real.

april-fools8gi8cocommunity

Read

March 21, 2026

Day 1: The Overnight Session That Built 8 Systems

One session. Four repos. 30+ agents deployed in parallel. 10,000+ lines shipped. KittenTTS voices for kids. A particle physics simulator. And a Karpathy interview dissection that rewrote our architecture.

dev-log8gentnick-osautoresearchkarpathy

Read

March 20, 2026

The case for local-first AI development

Why defaulting to local models and zero API keys is not just a privacy stance - it is a better development experience.

local-firstphilosophy

Read

March 15, 2026

Why we build a harness, not just an agent

Most AI coding tools ship a chatbot. We ship an evaluation harness that grades execution, not vibes. Here is why the harness matters more than the agent itself.

architecturebenchmarks

Read

March 12, 2026

Execution-graded benchmarks: methodology and results

Our benchmarks test whether code actually runs, not whether it looks right. A deep dive into how we score agent output with deterministic, reproducible grading.

benchmarksmethodology

Read

March 8, 2026

GLP-stage adaptation: building AI for neurodivergent children

How 8gent Jr adapts to gestalt language processing stages, supporting autistic and ADHD children with personalised interaction patterns.

accessibility8gent-jrneurodiversity

Read