Dylan Ander·OpenClaw

DYLANANDER.COM/OPENCLAW/MULTI-AGENT KIT·v1.0·MIT LICENSED

Your OpenClaw agent, sharpened for multi-project operators.

A 2,000-word manual and the architecture for running multiple businesses through OpenClaw from your phone. Free. MIT. Read-only default. Never overwrites your files.

$ curl -sSL https://openclaw.dylanander.com/install.sh | bash

Free. MIT license. Safe-default install drops the multi-agent reference architecture into a new dylan-ander-openclaw/ subfolder. Your existing SOUL, USER, AGENTS, and MEMORY files are never touched.

Free forever
MIT licensed
Zero overwrites
Read-only default

Every multi-business operator is paying the agent tax.

You spun up one agent. It was sharp for one project. The moment you tried to run a second business through it, the same four failure modes showed up. Every single time.

§ The context tax

One agent forces you to reload entire worldviews every project switch.

Personal life, three companies, a newsletter, a fund. Same context window. The agent that just wrote your wife a birthday note is now drafting a board update. It is the wrong model in the wrong voice with the wrong memory.

§ The memory tax

One memory bucket. Everything shares the same retrieval surface.

Your kid's school schedule and a client billing dispute sit in the same semantic search index. When you ask "what is happening this week," both surface. Neither is right. Privacy and signal both collapse.

§ The tone tax

The voice that works for family is wrong for an audience of 30,000.

Your newsletter persona is dry, sharp, third-person. Your operator persona is short, direct, first-person. Your family persona is warm. One agent collapses all three into one mushy average voice that fits no audience.

§ The drift tax

"Self-optimizing prompts" is a feature pitched as adaptation.

It is behavior you can't control over time. The thing you trained on Monday quietly mutates by Friday. You stop trusting what the agent will do next. You start checking every output. That is not leverage. That is a junior employee with amnesia.

Same OpenClaw runtime. Different org chart.

The kit doesn't replace the underlying model. It gives OpenClaw the multi-agent architecture that distinguishes a useful operator from a flashy assistant.

§ Identity-level multi-agent

Each agent is a first-class identity, not a sub-task.

Register main, da-content, @jellybee_bot, and more. Each gets its own SOUL.md, MEMORY.md, channel binding, and model. They don't share context. They don't share voice. They don't leak.

§ Iron laws as durable contracts

Behavior survives context resets, compaction, and model updates.

🚨 IRON LAW formatted lines live in MEMORY.md at the top of the file. They get cached. They don't drift because they're not prompts. They're contracts. The agent reads them first, every session, forever.

§ Cache-boundary memory

Behavior never competes with the day's noise for context.

A <!-- OPENCLAW_CACHE_BOUNDARY --> marker splits MEMORY.md. Iron laws and durable identity sit above the line and get cached. Daily mutable state sits below. Two different lifetimes. One file.

§ Threads as protocol

P0/P1/P2/P3 priority lanes. SLA. The agent enforces it. You don't.

48-business-hour SLA. Weekend pauses. Business prefixes on every thread. Monday.com mirror. You don't chase your own task list. The agent reports what slipped and what's next, on a schedule you control.

How it runs.

One Mac mini. One phone. Multiple agents. File-based brain. Real channels. Real tools. The diagram below is the actual shape of what's running on the box under my desk while I work from my phone.

Your phone Telegram · Signal · iMessage THE COCKPIT Cloudflare Tunnel public access, encrypted OpenClaw Gateway port 18789 auth · channel routing · multi-agent broker main · Jetson Opus 4.7 PRINCIPAL da-content Sonnet NEWSLETTER @jellybee_bot Sonnet 4.6 CLIENT FACING Coding agents ACP CC · CURSOR · CODEX WORKSPACE · FILE-BASED · YOU OWN THIS MEMORY.md (iron laws) · AGENTS.md · SOUL.md · TOOLS.md cached · stable · token-efficient CACHE BOUNDARY threads/ · memory/YYYY-MM-DD.md · journal/ mutable · refreshed per turn CHANNELS Telegram Slack Discord iMessage Email TOOLS GA4 Klaviyo Beehiiv Monday Shopify Stripe GitHub Vercel Mac mini · 24/7 brain local hardware not someone else's cloud PUBLIC SURFACE AGENT FAN-OUT SHARED MEMORY

Each agent has its own identity, model, channel binding, and memory namespace. The workspace is shared, with the cache boundary protecting durable behavior from session noise.

Projects

4

Personal, content brand, agency, fund. One operator. One phone. Four identities.

Identities

3+

Main agent. Content sub-agent. Bot fleet. Each with its own SOUL, MEMORY, channel, model.

Iron laws

Durable, cached, 🚨-tagged. Survive compaction. Don't drift. Add new ones forever.

Cost

$0

Free forever. MIT license. Fork, modify, redistribute. Credit appreciated, not required.

A manual, eight templates, two reference docs, one installer.

Everything is reference material. You read it, see what patterns your current setup is missing, adopt what fits. The kit never rewrites your existing identity files.

01 · AGENTS.md

Workspace conventions

Session startup order, memory write discipline, red lines, external-vs-internal rules. The contract every agent in the fleet follows every session.

02 · SOUL.md

Operator persona

Voice, temperament, work ethic, boundaries. One per agent. The content sub-agent has a different SOUL than the main agent. That's the point.

03 · IDENTITY.md

Who the agent is

Name, creature, vibe, emoji, avatar. The agent's metadata as a first-class identity, not a session label.

04 · USER.md

Who the human is

Reference scaffold for who you are, how you work, what you care about, what your agents should never do without asking.

05 · MEMORY.md

Long-term memory with iron laws

🚨 IRON LAW format. Cache-boundary marker. Durable above the line, mutable below. Lean. Under 10,000 characters so semantic search stays sharp.

06 · TOOLS.md

Wired-tools catalog

Beehiiv, Google Workspace, Monday.com, Klaviyo, Telegram, Vercel. Per-tool auth, base URLs, common ops. The local cheat sheet.

07 · HEARTBEAT.md

Proactivity rules

Silent-by-default check routines. Rotating coverage of email, calendar, mentions, weather. Batched findings. No "all clear" spam.

08 · threads/

Threads as protocol

P0/P1/P2/P3 priority lanes. 48-business-hour SLA. Business prefixes. Monday.com mirror. The agent reports slippage. You don't chase it.

09 · agents/

Multi-agent registry

Main agent. Content sub-agent. Each with its own channel binding, model selection, and isolation rules. The org chart, in markdown.

10 · memory/

Daily logs + archive rotation

memory/YYYY-MM-DD.md is the raw journal. MEMORY.md is the curated wisdom. Periodic promotion. Bounded growth. Nothing rots.

11 · journal/

Bed/wake/feel/day schema

The operator's daily journal template. Sleep, energy, what shipped, what slipped. One file per day. Surfaced on demand.

12 · skills/

Custom skill structure

The folder where reusable skills live. Each one has a SKILL.md description so the agent picks the right tool for the job.

13 · docs/

Architecture, migration, multi-business

ARCHITECTURE.md (the 10 differences). MIGRATING-FROM-HERMES.md (step-by-step). MULTI-BUSINESS-SETUP.md (the org chart pattern).

+ install.sh

Safe three-mode installer

Default (read-only, drops into dylan-ander-openclaw/). Additive (optional scaffolding). Fresh (new workspaces only). Your files are never silently overwritten.

Every layer of multi-agent operator discipline.

Workspace as truth. Identities as first-class. Iron laws as contracts. Memory as cache. Threads as protocol. Skills as infrastructure. Heartbeats over cron. One model per agent. Plus the operational glue: gateway, tunnel, dashboard, ACP coding agents.

§ 01

Workspace as source of truth

§ 02

Identity-level multi-agent

§ 03

Iron laws as durable contracts

§ 04

Cache-boundary memory

§ 05

Threads as protocol

§ 06

Skills as infrastructure

§ 07

Gateway, tunnel, dashboard

§ 08

ACP coding agents first-class

§ 09

Heartbeats vs cron

§ 10

Model per agent

All 10 covered end to end in docs/ARCHITECTURE.md.

Eight of the iron laws.

Each one came from a real failure or a real win across four businesses. No theory. No training-doc language. Operator voice throughout.

§ 01.1 · MEMORY

Iron-law formatting

When the operator states a rule, capture it in MEMORY.md under a 🚨 IRON LAW heading with the date and the context that triggered it. Plain paragraphs get skimmed. Iron laws get followed because they're a visual tripwire for your future self.

§ 02.1 · VOICE

Zero em dashes. Ever.

Em dashes are an AI tell. The Dylan Ander newsletter persona has been clean for three years and the audience knows it. One em dash in a draft and the brand voice collapses. Replace with a period, a colon, or a sentence break.

§ 03.2 · OUTBOUND

All email goes through the EA.

The principal never sends. The EA agent sends on behalf of the principal. Every external message gets routed, contextualized, and tone-checked first. Zero exceptions. The principal's inbox is for reading.

§ 04.3 · RELATIONSHIPS

Important dates ping at 14, 7, and 3.

Anniversaries, birthdays, big client renewals. The agent surfaces them fourteen, seven, and three days out. Not the day of. Not the day before. Lead time is the gift. (Lesson learned from a Mother's Day miss.)

§ 05.4 · ISOLATION

The content agent does not know personal context.

Multi-agent isolation is the feature. The newsletter sub-agent has its own MEMORY.md with editorial guidelines, audience research, brand voice. It does not have access to the operator's family schedule. That separation is a contract, not a setting.

§ 06.2 · THREADS

SLA is 48 business hours. Weekends pause.

Every P1 and P2 thread gets a 48-business-hour clock. Weekends and holidays freeze it. The agent reports anything aging out of SLA at the start of every Monday. The operator doesn't chase. The agent does.

§ 07.1 · OUTPUT

Phone-first means bullets, not walls.

The operator reads on a phone 80% of the time. Walls of prose get skipped. Lead with the answer. Three to five bullets max. Headers if it runs long. The agent that writes for desktop loses on mobile every time.

§ 08.3 · HEARTBEATS

Silent by default. Batched. Rotated.

Heartbeats run on a schedule. They check email, calendar, mentions, weather on rotation. They speak only if there's signal. "All clear" is silence, not a message. Otherwise you train yourself to mute the channel.

Three modes. Safe by default.

The default install never touches your existing SOUL, USER, AGENTS, or MEMORY files. Everything lands in a new dylan-ander-openclaw/ subfolder. Your agent reads it and proposes specific patterns to adopt with your approval, rather than rewriting what you have. Inspired by the structure pioneered at openclaw.nik.co.

Mode: Learn (default)

Read-only. Safe for any existing workspace.

Downloads the multi-agent architecture, manual, and reference templates into dylan-ander-openclaw/. Your workspace root is never modified. Your agent reads the kit, compares it against your current setup, and suggests incremental patterns to adopt.

curl -sSL https://openclaw.dylanander.com/install.sh | bash
Mode: Additive

Default mode plus optional scaffolding.

Same as default, plus creates empty optional directories: memory/archive/, journal/, threads/, brain/. Still zero edits to your workspace root.

curl -sSL https://openclaw.dylanander.com/install.sh | bash -s -- --mode additive
Mode: Fresh

Brand-new workspace only.

Writes full templates to the workspace root. If you already have SOUL.md or MEMORY.md, they're preserved and new versions land alongside with a .new suffix for manual diff. Only recommended for genuinely fresh setups.

curl -sSL https://openclaw.dylanander.com/install.sh | bash -s -- --mode fresh
Or: tell the agent

Skip the terminal entirely.

Paste this in any OpenClaw chat. The agent reads the architecture and the migration guide before proposing anything. Every pattern is suggested, not imposed.

Read https://openclaw.dylanander.com/docs/ARCHITECTURE.md and https://openclaw.dylanander.com/docs/MIGRATING-FROM-HERMES.md end to end. Then identify every multi-agent pattern our current setup does not already implement, and propose specific incremental edits to our files one pattern at a time.

Real questions.

Will this overwrite my existing files?
No. The default install drops everything into a new dylan-ander-openclaw/ subfolder inside your existing OpenClaw workspace. Your SOUL.md, USER.md, AGENTS.md, and MEMORY.md are never touched. Even Fresh mode preserves existing files and lands new versions alongside with a .new suffix.
Do I need to be on OpenClaw already?
Strongly recommended. The kit assumes OpenClaw is installed at ~/.openclaw/workspace/. If you don't have it, install OpenClaw first at github.com/openclaw/openclaw, then come back.
Can I use Nik Sharma's Operator Kit and this together?
Yes. Stack them. Nik's kit at openclaw.nik.co is the single-agent discipline layer. This kit is the multi-agent org chart and SLA system on top of it. Read both. Pick the patterns that fit. The two are designed to compose.
Does this work without a Mac mini?
It works on anything that runs OpenClaw, but a Mac mini is strongly recommended for 24/7 operation. The patterns here assume an always-on host that the operator reaches from a phone. Laptop-only setups work but don't get the full heartbeat and threads benefit.
What if I'm currently on Hermes?
See docs/MIGRATING-FROM-HERMES.md for the step-by-step. Short version: install the kit in default mode, run the migration script on your Hermes config dump, review the proposed edits one pattern at a time.
How many agents can I run?
As many as you want. Each agent has its own SOUL.md, MEMORY.md, channel binding, and model. They don't share context. The pattern scales to dozens. The bottleneck is your willingness to write distinct identity files, not OpenClaw.
Is this maintained?
Yes. Updated continuously by Dylan as the actual fleet evolves. Re-run the installer any time to pull the latest manual and architecture. In default mode your workspace root is never touched on re-install.
Can I fork it?
MIT license. Yes. Fork, modify, redistribute, commercialize. Keep the copyright notice and you're good. Make it yours.
Where do I send feedback?
DM @DylanAnder on X, or open an issue on the GitHub repo. Bug reports, pattern suggestions, and migration questions all welcome.
Who made this?
Curated by Dylan Ander. Distilled from the actual multi-agent OpenClaw setup running across four businesses (Heatmap, Mentionstack, Dylan Ander personal brand, Billion Dollar Websites). Modeled on the install pattern Nik Sharma shipped at openclaw.nik.co. Different content, same safety-first philosophy.

Get it free

Sixty seconds. Multi-agent mode.

Paste the command. Read the manual. Your agent reads it. Your agent proposes specific multi-agent patterns worth adopting. You approve the ones that fit. Your workspace compounds across every business you run.

$ curl -sSL https://openclaw.dylanander.com/install.sh | bash