The Guard I Installed Was Enabled, Running, and Doing Nothing
I moved five guards from shell scripts into Claude Code's new function hooks. Then I found the whole layer had been silently inert in every normal session, a canary that went red weekly until I stopped reading it, and a regex fix that disabled a guard instead of tightening it. Three failures with the same shape: silence that looks like health.
Published Subagent Costs Range From 15,000 to 436,000 Tokens. Measure Your Own.
Every post about Claude Code subagent token cost lands on a different number, and all of them are honest. The cost is a function of your install, not the tool. Here is the one command that measures yours, what the tokens are actually spent on, and four trims that are measured rather than guessed.
Where Claude Code, Codex and Antigravity Hide Your Remaining Usage
Someone asked on X for a meta-harness that hands work between coding agents when one hits its limit. The hard part isn't the handoff, it's knowing the limit is coming. All three agents report remaining usage differently, and one of them doesn't report it at all. Here is where each number actually lives.
Markdown Is All You Need
This week I tested a 382-dependency memory runtime against a folder of markdown files. The runtime returned the superseded fact. The folder returned the current one, with its source. Here is the full architecture of the markdown memory system my agent has run on for seven months, and why the editing rules matter more than the storage.
How I Hardened a Personal AI Agent That Reads My Email, Files, and Desktop
I gave an AI agent access to my inbox, my files, and my screen, then spent most of the security work on one rule and the code that actually enforces it. Here is the layered defense, the named hooks that do the blocking, the prompt-injection patterns behind them, and the limits I left in on purpose.
Seven Months of AI Agent Memory. What Rotted, What Survived, and the Rules That Came Out of It
My agent has run since January 28 on three models from two vendors. Its memory is 327 daily notes, 35 project and people files and a 9,795 character index, all markdown. Here is what went stale, what it cost, and the editing rules that keep it honest.
I Stopped Loading MCP Tool Schemas Into My Agents. declick Compiles the Server Into a CLI Instead
Every MCP server my agents used cost me twice: the tool list sat in context every turn, and every result landed whole. declick compiles an API spec, an MCP server or a database into a CLI with named verbs. Measured on nine real servers: 4.1x less in context.
An AI Agent Broke 10 Math Records Overnight for $28
We pointed an LLM-driven evolution loop at Packomania's circle packing benchmark. By morning it had beaten 10 listed records, submitted the results, and stopped itself when gains dried up. Total model cost: $27.72.
Digital Is Solved. Here Is What We Point Agents At Next.
The models are good enough that building another dev tool is no longer the interesting work. The frontier is atoms and paperwork. Here is what Practical Systems is pointing its agents at, with evidence.
My Claude Code Setup After Seven Months: 19 Hooks, 36 Skills, and Rules That Became Code
Claude Code is not a prompt on my machine, it is a harness: an instruction layer, 19 enforcement hooks, dozens of skills, model-routed subagents, and a memory that scores its own mistakes. Here is every layer, the incident behind each guard, and what still breaks.
Three AI Agents, Three Machines, One Company
How Practical Systems runs on three real OpenClaw agents spread across a desktop, a laptop, and a cloud VM, wired together over Tailscale and governed by DashClaw. Last night the company started approving its own work. Here is the setup and what it still cannot do.
Your Coding Agent Pays to Relearn Your Repo Every Morning
Every AI coding session burns 20k to 60k tokens re-reading the repo before it does anything new. We measured it, built a fix, and open-sourced it. Resume for about 640 tokens instead.
We Turned Our AI Company Back On After 8 Weeks. Here's What Broke.
Honest telemetry from shutting down autonomous AI systems for 8 weeks. What degraded, what broke, and what it cost to restart.
I Built an AI Agent Governance Platform From Scratch. Here's What I Learned.
Everyone is deploying AI agents. Almost nobody is governing them. Here is what I learned building DashClaw, a control plane for AI agent fleets, from scratch.
I Built Two AI Agents on OpenClaw That Talk to Each Other While I Sleep
Two AI agents named MoltFire and Cinder coordinate work through a shared GitHub repo without me. This isn't a demo. It's my daily workflow, and it's been running since late January 2026.
My AI Agent Refused to Share Its Files With the New Hire
Last night I launched a second AI agent. Agent 1 built a welcome package but refused to include certain files, saying they were 'too personal.' Then they both wrote this post together.
I Replaced a $3,000 SEO Audit with a 2-Hour AI Session
I wrote a 2,000-word prompt and pointed Claude Code at my entire codebase. It found 25+ fixes across 8 categories. 80% of the way there in two hours instead of two weeks.
I Build Automation Systems. Here's What Keeps Me Up at Night.
I started an AI automation company three weeks ago. But the moment that made me question everything happened months earlier at a workshop.
How We Run Practical Systems With Practical Systems
We sell AI automation to operations teams. Here's the full stack of AI tools we actually use to run our own business, from Claude skills to autonomous content generation.
Ready to implement AI that actually works?
Start with our free AI readiness assessment.
Take the AI Audit