Reasonbee — ЯB. hexagon mark Reasonbee Comms

Reasonbee Comms — AI Blog

AI Blog

Build logs, links and the occasional essay — hands-on notes from building AI. Posted here first.

Maker's dashboard // cached — live metrics unavailable at build
66.79B
agent tokens (cache incl.)
Claude 59B · Codex 7.79B
from 14 Apr 2026 → 21 Jul 2026
3,810hrs
Hours in Claude Code
8h/day since 1 Apr 2025 (UTC)
4,334
commits
cached snapshot
91
repositories
owned, non-fork
8
languages
TypeScript · JavaScript · Python · HTML · Shell · Astro · C++ · GDScript
Claude Code + Codex
main go-to
declared
8
agents I run
Claude Code · Codex · Gemini · DeepSeek · GLM 5.2 · MiniMax · Kimi · Mercury
Vercel · AWS
deployed to
declared
The feed newest first · 4 posts RSS ↗
type tags
essay

Ten million tokens for one conversation

The first version worked — expensively. How we took a recruitment agent from ten million tokens a conversation down to something a business can actually run.

#agents#tokens#optimisation

essay

The bottleneck in building agents is human communication

The market thinks the hard part of AI is the AI. Build agents for real businesses and you learn the truth: the bottleneck is getting what's in people's heads into a shape an agent can use.

#agents#knowledge#okf

essay

I rented a stack of H200s and ran GLM-5.2 myself. The small models had already won.

I rented cloud GPUs to run open AI models — from a 27B that fits a second-hand card to a 743B giant — and measured what actually holds up for a privacy-first small business. Plus the one conclusion that caught me out: renting a cloud server isn't privacy.

#local-ai#self-hosting#open-models#privacy#glm#vllm#runpod

essay

My Son's Teacher Told Him Not to Use AI

Schools are framing AI as something to fear. They've got it backwards — it's the best learning tool any of these kids will ever have.

#learning#education#ai-literacy#curiosity