AgentRoot

What it looks like

Say it once. It stays said.

State a rule or a fact in plain language. The assistant records it and tells you what it wrote. What it only guessed waits for your yes.

Your rules, every time.

Set the house rules once. Weeks and dozens of chats later, in another agent, they still hold.

Case 1illustrative

Cursor · Sep 29

YouHouse rules for this project: every task gets an owner and a due date. Never pick an owner yourself — ask me. Status updates list risks first.

AgentRemembered: "Every task has an owner and a due date."
Remembered: "Never assign an owner without asking."
Remembered: "Status updates list risks first."

… four weeks and thirty chats later …

Codex · Oct 27

YouTurn today's meeting notes into tasks.

Agent(reads Memory first) 9 tasks, each with an owner and a due date.
2 have no owner in the notes: "Update pricing page" and "Renew SSL certificate". Who takes them? I didn't guess.

YouMei, both.

AgentDone. Status, risks first: the SSL certificate expires Nov 1.

It catches what you'd miss.

Something said in passing two weeks ago stops a mistake today — and it shows you where it came from.

Case 2illustrative

Cursor · Oct 13

YouThe client moved the beta to Nov 3. And nothing goes to customers before Legal signs off — Dana signs.

AgentUpdated: "The beta is on Nov 3." (was Oct 20)
Remembered: "Nothing goes to customers before Legal signs off."
Remembered: "Dana signs off for Legal."

… two weeks later …

Claude Code · Oct 27

YouPlan this week's release.

Agent(reads Memory first) Six steps, working back from Nov 3.
⚠ Step 4 sends the build to two pilot customers. On Oct 13 you said: "nothing goes to customers before Legal signs off." I moved Dana's sign-off ahead of it.
Every plan you approved kept a day of buffer before a deadline. Keep that as a rule?

YouYes.

Nothing above needed a file edit. The exact wording depends on your model.

Three things, kept apart

What's true, how you work, who it helps.

Most agent chats forget everything between sessions. Prompt dumps and hand-edited YAML try to fix that, and become someone's maintenance burden. AgentRoot stores each concern in its own place, as flat files you can read.

What's true about this project

Memory

Short one-sentence facts. What you say is active at once; you can withdraw any of them.

How we usually do this kind of work

Skills

Procedures, kept only after you confirm them. A scored run that does worse sends one back.

Who the assistant is helping

Role

Identity and lens, not a step-by-step manual. General by default.

It learns how you work

Do it once. Next time, it already knows how.

When work repeats, the assistant offers to keep the way it was done — the steps, and the script if it wrote one. Nothing is kept until you say yes.

A procedure, learned.

Done once in Claude Code, kept as a Skill, run the same way weeks later in Codex.

Case 3illustrative

Claude Code · Oct 6

YouTurn these 14 vendor status reports into one tracker: vendor, milestone, due date, status, risk.

AgentDone: tracker.xlsx, 14 rows. 2 need your review — Vendor F gave no date; Vendor K's status contradicts its last report.
I wrote a script for the table part. Save the procedure and the script as a Skill, so next time it runs the same way?

YouYes.

… every Monday after that …

Codex · Oct 27

YouVendor reports are in.

AgentUsing Skill: vendor-tracker (you confirmed it Oct 6). 16 reports → tracker, same columns, same checks. 1 needs your review: Vendor B's due date slipped three weeks.

Only what repeatsIt offers a Skill for work that came back, or for batch work where it wrote a script the first time. Whether two runs were the same kind of work is read from what they did, not from the agent's guess.
Worse goes backOn scored work, a Skill that makes results worse returns to proposed.
Stale gets retiredRules and Skills nobody uses any more are flagged, and you decide whether they go.

How capability stacks

Climb when you need to. Nothing breaks if you don't.

Each level sits on the one below. A missing level weakens only itself — no MCP and no Score still leave builtins and Memory useful.

T5
Trace + ScoreScore each run against a signed Answer key
T4
MCP toolsOptional live tools, re-discovered each run
T3
Confirmed SkillsRepeatable procedures you explicitly approved
T2
MemoryProject facts persist across sessions
T1
Builtins + work/Research, Analyze, Brainstorm, Report out of the box

Quick start

Two commands, then open the folder.

Requires Python 3.11+. Install the agentzero command once — no clone needed.

# install once (pipx install works the same)
uv tool install git+https://github.com/jesselei1102-hue/agentzero

# make a workspace for your tool: cursor | codex | claude
agentzero init ~/AgentZero-Workspace --harness claude

Give tasks in plain language.No new commands to learn for day-to-day work.
State facts as you go."We use millimetres." Say "forget that" to withdraw one.
Answer what's waiting../a0 memory review lists it as sentences.
Hand over references.Standards and templates become Knowledge later chats can use.

Design principles

Conventions, not a new runtime.

Convention over runtime

Adapts onto the harness you have. The agent loop stays in Cursor, Codex or Claude Code.

Confirm before keep

Proposed until you say yes. Reuse must not quietly degrade.

Operator-visible files

Flat YAML and Markdown you can audit. No hidden control plane.

Domain-agnostic core

The first proof is not the product boundary.

Tiers degrade independently

Skip a level and only that level is gone.

Not a new agent runtime. The LLM loop stays where it is.

Not an SDK-first product. Files and chat are the surface; src/ is there if you want to dig.