Skip to content
Shijie XieSHIJIE / KEEP ROAMING
GitHub

NOTES FROM SHIPPING WITH CODING AGENTS

Hand the work to agents.
Make the decisions.

Been through every shift in AI coding: prompts, context, harness.
This is where I keep the tools I built and the notes from each.

Shijie Xie
Hi, I'm Shijie Xie.

I work on agents at MiniMax, currently on MiniMax Code Agent, and I write down what I learn along the way.

About me
Some tools, some notes, and a few ideas still in progress.

HOW THE WORKFLOW CHANGED

From more agents
to one agent that proves its work.

I started by running more agents at once. What helped was giving one agent the whole job and a way to check it.

Chapter 1 / Keeping the method

dev-skills

Decide what you want,
then let the agent run with it.

From requirement to a verified, mergeable PR: a set of Skills for coding agents. First, give the repository a verification environment the agent can start, operate and observe on its own. After that, one agent takes each requirement from start to finish and proves the result in the running app, and a model from another family verifies it independently. You step in at three points: confirm the decisions, confirm the acceptance criteria, and read the decision list before you merge.

Open the field kit

Chapter 2 / The first tool

Agent Lord

Before more agents,
get one agent right.

Where this workflow started: a tool for dispatching coding agents from one main session. Written in August, it hands tasks to Claude Code, Codex and MCode, lets them work in parallel and review each other, and records every step. Using it showed me the bottleneck wasn't the number of agents: they couldn't run the product to prove their work, and decisions I'd already made got lost across handoffs. Since late September, new work goes through dev-skills instead.

See what it did
The first stretch of the road
AGENT LORD / OBSERVER
Agent Lord Observer showing tasks, tool activity and review results
One window, every step.Interface demo · synthetic data
A record of every step it ran.

Chapter 3 / Writing it down

Writing

Go back over what I built,
and think it through.

Specific problems from building with AI, what I did about them, and the calls I made. English versions are listed here as they're written.

All English writing (1)
SELECTED WRITING

Agent workflow

I stopped babysitting my coding agents. Here's the workflow.

One eight-module job took 14 hours, and for about 76% of it only one CLI was working. What I learned from OpenAI, Anthropic and Lauren Tan, and the workflow and three skills I use now.

Chapter 4 / About

One question throughout

The tools keep changing.
The question hasn't.

From developer productivity to code intelligence to AI coding, I've kept asking the same thing: how do developers get more done?

More about me

A FEW STOPS ALONG THE WAY

Build something,
then keep going.

  1. 01
    Developer productivityKuaishou · 2021 — 2023
  2. 02
    Code intelligenceLi Auto · 2023 — 2025
  3. 03
    AI codingXiaohongshu · 2025 — 2026
  4. 04
    AgentsMiniMax · 2026 — now
Shijie Xie / AI Roam

CONTRIBUTING IN THE OPEN

Open-source contributions

Problems I hit in real work, fixed in the tools we share.

MiniMax Code / PR #280Merged ·

A slowdown, or a noisy machine?

Paired re-runs in performance CI, to tell a real regression apart from runner noise and from too little evidence.

The performance budgets stay the same. Routine checks warn on persistent noise; the full performance check still needs a clear result.

See the pull request

RECENT NOTES

Recent updates

THERE IS ALWAYS ANOTHER PAGE

Stay curious.
See you on the next one.

New English writing goes on X. Code lives on GitHub.