HarrisonSec — Long-form writing on agent runtimes, distributed systems, and security

Jev returns a confidence with every decision.

Every tutorial on hiding a message in a photo teaches the same method: least-significant-bit replacement.

WaxSeal: Robust Image Watermarking for Secret Delivery The problem You need to hand a credential to a named colleague.

Pi contains no MCP support.

Codex made Item a wire type.

Pi’s reputation rests on one number: a system prompt small enough to quote.

Two API settings.

You can attach GDB to a stock Ubuntu 6.8 kernel running in QEMU and step through it with full symbols — no custom build, no kernel compile.

Set a breakpoint by name in a real-mode bootloader and it works.

Everything you write at work stands on a tower of things that were already set up for you: a heap, a stack, threads, a filesystem, a C library, an operating system underneath all of it.

Here is a bug report that keeps recurring, in different words, every time a team moves to ARM: “the same code works on our Intel CI and on the developers’ older MacBooks, but it corrupts data / deadlocks / returns impossible values on Graviton.”

Read enough x86 and you start to narrate it wrong in your head: “ja .target — jump if the first operand is above the second.”

“User space” and “kernel space” sound like operating-system inventions.

There are two common ways to attach an eBPF program to a kernel function: kprobe and fentry.

top says both threads are at 100% CPU.

You can rebuild an ARM64 Linux kernel, boot it, and single-step through start_kernel in GDB without owning a single piece of ARM hardware.

Codex ships 43,591 lines of sandboxing code across four platform backends.

Claude Code’s agent loop is 1,729 lines of TypeScript.

Sixty-five bytes is not a model.

There is a version of a bug that is more frustrating than an unfixed one.

I spent two days on a bug where a screen-capture stream never came back after the display went to sleep.

Here is a failure that survives every reasonable thing you can check.

An agent is asked to send a payment confirmation email.

A team I was helping watched their model bill jump to several times its usual size in a single month.

An agent I was reviewing was asked to rename a function across a repository.

There is one failure mode I have learned to fear more than a missing check.

I have now written this same sentence three times in three pieces, so it is time to write the sentence underneath it.

I keep writing variations of the same sentence.

I sat down to benchmark a tool and ended up with a map of a wall.

Higgsfield shipped a Minecraft “prompt-to-build” feature: a mod that drops a “Supercomputer” block into your world, takes a free-text prompt, and generates a structure in-world a minute later.

In April 2026, Anthropic published a blog post called “The advisor strategy: Give agents an intelligence boost”, naming a pattern they had been A/B-testing in production: a cheaper model runs the agent loop end-to-end, an expensive model is consulted only…

The prior post in this series, Agent Retrieval Is a Cost Curve Problem, argued that a viable LLM-symbol-graph would need to satisfy six specific conditions — and that no existing tool had hit all six.

Two weeks ago CodeGraph hit GitHub trending — tree-sitter + SQLite/FTS5 + MCP for Claude Code, 19k+ stars in a week.

This post is one half of a pair.

There’s a popular interview question making the rounds: “Why doesn’t Claude Code use RAG to retrieve code?

It’s 3am.

The Setup I was optimizing an AI agent for a production system — a creator agent that handles user requests like “make this character fiercer” or “rename this entity.”

The Problem Nobody Talks About AI can write code, generate content, analyze data, design systems, and manage projects.

Most discussions of TypeScript/Node vs Go concurrency stop at the surface: Node is async, Go is threaded.

The model is roughly 10% of what makes an AI agent work in production.

Episode Summary Building an AI agent that works is easy.

Long-form podcast version of the 90% Problem thesis.

Your Agent Works Great.

You’re debugging a gnarly race condition.

Every AI coding agent runs the same core pattern: send context to an LLM, get back text and tool calls, execute tools, feed results back, repeat.

This is Part 4 of our Claude Code Architecture Deep Dive series.

This is Part 3 of our Claude Code Architecture Deep Dive series.

This is Part 2 of our Claude Code Architecture Deep Dive series.

Adding AI API calls to an existing backend is where most teams’ observability and billing instincts break.

Updated: This analysis has been superseded by Part 4: Why It Uses Markdown Files Instead of Vector DBs — a deeper first-principles tradeoff analysis.