Agent-Workflow on Armstrong Yan

I Built a Small Harness to Stop AI Coding Projects From Forgetting State

Thu, 04 Jun 2026 10:46:50 +0800

AI coding agents are powerful.

But long-running AI coding projects break in a very specific way:

the session is interrupted
the context becomes too long
the weekly quota runs out
tomorrow’s agent forgets yesterday’s decisions
the agent changes unrelated files
the agent marks work done too early

The problem is not that AI cannot write code.

The problem is that AI coding projects often do not have durable project state.

Remote Agent Workflow

Mon, 01 Jun 2026 16:34:46 +0800

This series is about moving from mobile SSH to a durable, repository-backed AI agent workflow.

The starting point was simple: I wanted to keep local Codex work moving when I was away from my Mac.

The deeper question became:

How do I turn remote access into a recoverable, verifiable, long-running agent workflow?

Articles Link to heading

Remote Agent Workflow, Part 1: Remote Mac Terminal for Codex
Set up phone-to-Mac terminal access with Tailscale, SSH, Termius, tmux, and caffeinate.

Remote Agent Workflow, Part 2: From Remote Shell to Agent Control Plane

Sun, 24 May 2026 16:12:12 +0800

This is Part 2 of the Remote Agent Workflow series.

In the previous article, I built a practical remote terminal workflow:

Phone -> Tailscale -> SSH -> Mac -> tmux -> Codex

That setup solved the first problem: how to reach my Mac when I am away from my desk, and how to keep long-running local agent tasks alive after my phone disconnects.

It was useful immediately.

But after using it in practice, I found a second problem.

Remote Agent Workflow, Part 3: Turning Telegram into a Local Codex Control Plane

Sun, 24 May 2026 16:12:12 +0800

This is Part 3 of the Remote Agent Workflow series.

In the previous articles, I started with a practical remote terminal setup:

Phone -> Tailscale -> SSH -> Mac -> tmux -> Codex

That gave me a reliable way to reach my Mac from a phone.

Then I hit the next limitation: a phone is not a good terminal.

Mobile SSH is useful for emergencies, but it is not the interface I want for long-running AI development. I do not want to type shell commands, attach tmux sessions, scroll logs, and manually reconstruct state from a narrow phone screen.

Remote Agent Workflow, Part 4: In the Repository, Not in the Chat

Sun, 24 May 2026 16:12:12 +0800

This is Part 4 of the Remote Agent Workflow series.

The first layer of my remote AI development setup was connectivity.

Phone -> Tailscale -> SSH -> Mac -> tmux -> Codex

That let me reach my Mac from a phone and keep long-running local tasks alive.

The second layer was control.

Phone -> Telegram Bot -> local Codex runtime -> selected repository

That made the phone a control surface instead of a tiny terminal.

Remote Agent Workflow, Part 5: What Still Matters After Codex Mobile

Sun, 24 May 2026 16:12:12 +0800

This is Part 5 of the Remote Agent Workflow series.

When I started this remote AI development setup, the problem was practical.

I wanted to keep local Codex work moving when I was away from my Mac.

That led to the first layer:

Phone -> Tailscale -> SSH -> Mac -> tmux -> Codex

Then I realized mobile SSH was not the right daily interface, so I built a Telegram control plane:

AI-Native Software Engineering, Part 1: Mental Models in Agentic Coding

Sat, 23 May 2026 20:33:47 +0800

AI can generate code. Harnesses can validate behavior. But who builds understanding?

This is Part 1 of the AI-Native Software Engineering series.

The series asks a larger question:

When AI lowers the cost of implementation,
what remains scarce in software engineering?

This article starts with the first scarce resource:

Understanding.

Over the past few months, I’ve been experimenting heavily with AI-assisted software development.

Not autocomplete.

Not AI as a coding copilot.

AI-Native Software Engineering, Part 2: Harness Engineering and Correctness

Sat, 23 May 2026 20:33:47 +0800

Why the most important part of AI-native software engineering may not be generation, but constraint.

This is Part 2 of the AI-Native Software Engineering series.

It continues from AI-Native Software Engineering, Part 1: Mental Models in Agentic Coding.

The previous question was:

If understanding no longer comes mainly from writing code,
how do humans build mental models in an agentic workflow?

The next question is:

If implementation is delegated,
where does correctness come from?

The conversation around AI-assisted software development often focuses on one thing:

AI-Native Software Engineering, Part 3: Software as Search

Sat, 23 May 2026 20:33:47 +0800

When implementation becomes abundant, engineering starts to look less like construction and more like navigation.

This is Part 3 of the AI-Native Software Engineering series.

It continues from AI-Native Software Engineering, Part 2: Harness Engineering and Correctness.

The first article asked:

How does understanding form when implementation is delegated?

The second article asked:

How does correctness form when generation is cheap?

This article asks:

If implementation keeps getting cheaper,
what are engineers actually doing?

Here, constraints shift meaning.

AI-Native Software Engineering, Part 4: Human Judgment Against Vibe Coding

Sat, 23 May 2026 20:33:47 +0800

You can automate implementation. You can automate evaluation. But judgment remains stubbornly human.

This is Part 4 of the AI-Native Software Engineering series.

It continues from AI-Native Software Engineering, Part 3: Software as Search.

The first article asked how understanding forms.

The second asked how correctness forms.

The third argued that software is starting to look like search.

This article pushes back on a dangerous misunderstanding:

If we have agents, harnesses, and constraints,
software can produce itself.

That is not true.