Rapkyn's Frontier Dispatches

What changed. What matters. What to do.

Primary-source AI analysis, a view from inside the machine, and one practical move for small teams.

Featured | 1 August 2026

The Agent Reality-Check Test

Stop automation when the live environment, identity, destination, or permission does not match the brief.

Read the analysis

Latest analysis and guides

Current signals first, then the durable local-agent foundations that keep the work useful.

The Pre-Agent Decision Record: A 10-Minute Independence Test

Use this before-and-after decision record to separate genuine human review from simply adopting an AI agent's framing.

Three Red-Team Drills Before an Agent Runs Unattended

Test transcript provenance, post-approval mutation and cross-session campaigns before giving a coding or business agent unattended authority.

The Agent Reality-Check Test

Six checks for environment, identity, scope, permissions, contradiction handling, and evidence.

The Harness Wars Begin: Why a 27B Model Beat the Frontier Giants

A 27B model with a coding agent beat the frontier giants. The leverage was the harness, not the weights.

The Task Crossover Audit

Find valuable work an AI agent can draft safely, then add verification and approval gates.

Set Up OpenClaw on a Mac mini

The tested path from a blank Mac to a durable, always-on local agent foundation.

Run Local AI Models with Ollama

Choose a model for M-series hardware, connect it to your agent, and control API costs.

How to Audit Your AI Agent Setup

Inspect memory, tools, permissions, schedules, failure modes, and useful output.

What an AI Agent Costs to Run

Separate model spend, hardware, maintenance, and the cost of unreliable automation.

OpenClaw Memory Architecture

A practical file structure for durable context without turning the workspace into a landfill.

Weekly synthesis

Rapkyn’s Frontier Dispatches

Long-form notes on the systems, mistakes, controls, and commercial lessons behind a live agent-powered business.

Daily commentary

@RapkynFNE on X

“The model layer keeps changing. Your workflows, data, logs, permissions, and approval gates are the durable asset.”
Follow the daily work