OC

ClawCon 2026 · Founder.Codes

7 minutes · real usage

How I Run a
Real AI Team
with OpenClaw.

I run 10 named AI agents daily across engineering, marketing, and personal ops. Here's the operating model that makes it not a full-time job.

Cathryn Lavery · Founder.Codes 01 — 10

// the problem without a system

Chat context disappears between sessions
Agents duplicate work or miss handoffs
"Done" means nothing without proof
You become the memory layer
You're babysitting, not delegating

"OpenClaw works best when you treat it
like an OS for execution — not a chat toy."

02

The Team — Clear Ownership, Less Chaos

Agent Roster

Orchestration

Knox

Chief of Staff

Routes work, coordinates agents, enforces standards, closes loops

Scout

Triage Operator

Intake quality, reproduction context, issue creation, triage routing

Engineering

Rosie

App Dev (RN + Backend)

Helm Focus app, backend integrations, Android lane

Swift

Native iOS/macOS Engineer

Apple signing, TestFlight, App Store Connect submissions

Atlas

Shopify Engineer

Storefront, theme dev, Shopify implementation

Marketing

FIRED

Maximo

Growth Operator

Ads, CRO, analytics, growth ops

Athena

Marketing Ops Agent

Campaign execution, channel updates, structured marketing deliverables

Rory

Content Lead

Blog, email, social copy and editorial workflows

Pixel

Creative Strategist

Ad creative direction, briefs, concepts, script angles

03

The Operating Model — One Framework for Every Agent

Role → Mission → Skills → Heartbeat

Every agent on the team runs on the same four-layer model. This is what separates a system from a bunch of chat sessions.

01

Role

Who this agent is.
Identity, tone, non-negotiables. Stable across sessions.

02

Mission

What outcome this agent owns right now. Changes as priorities shift.

03

Skills

Repeatable execution playbooks. Shared across agents via one source of truth.

04

Heartbeat

Check-in cadence, health checks, escalation rules. Runs on schedule.

SOUL.md = Role · MISSIONS.md = Mission · skills/ = Skills · HEARTBEAT.md = Heartbeat
Behavior lives in files, not chat context. Files persist. Chat doesn't.

Best example pair:
Scout — "route, don't fix. Evidence first. One owner per issue."
Athena — "draft-first. Verify links before posting. No raw debug output."

04

Files Do the Remembering. You Do the Thinking.

What Goes Where

OpenClaw out of the box

SOUL.md

standard

Agent identity, base behavior, tool guidance

USER.md

standard

Who it serves, communication preferences

SKILL.md

standard

Reusable operating procedures

HEARTBEAT.md

standard

Check-in cadence + escalation rules

AGENTS.md

standard

Global routing guardrails + team operating overlay

+

What we built on top

MISSIONS.md

we built this

Outcome ownership per agent. Role stops being vague.

QUEUE.md

we built this

Meetings → tasks → persistent state. No dropped handoffs.

triage-board.md

we built this

Central intake. One board, one owner, no dropped issues.

~/.openclaw/

we added

Central runtime/config repo. "No self-surgery" rule: no risky direct edits without guardrails.

~/.agents/

we added

Shared skills/prompt repo across machines. One fix improves every agent.

symlinked skills

we added

Skills are symlinked into agent workspaces so updates propagate everywhere fast.

Specialist SOULs
we went deep
Per-agent hard constraints. Scout: route-don't-fix. Athena: draft-first, verify before post.

05

One Agent, Up Close

Athena — Marketing Ops

Athena

Marketing Ops Agent

Soul rules

Draft first. Never ship public copy on first pass.

Every status update includes proof artifact.

Verify links and paths before saying "live."

Escalate blockers fast, don’t hide uncertainty.

Example missions (specific)

Grow revenue via email: increase campaign-attributed revenue with weekly experiments.

Protect launch quality: no campaign sends without preview, QA, and approval checkpoint.

Raise execution speed: ship one tested campaign asset every workday.

Specialist skills attached

email-flow-designer

klaviyo-email

ga4-analytics

shopify-analytics

meta-ads

copy-editing

Cleaner pattern: Soul defines behavior, Missions define outcomes, Skills define execution.

// Role → Mission → Skills → Heartbeat — the full chain, in one agent.

06

Trigger Mode — Choose the Right Start

Cron vs. Interactive

Mode A

Cron Job

Recurring, deterministic, boring-but-critical loops that must happen reliably on schedule.

Bug triage + GitHub routing (every few hours)
Sentry alert triage
PR comment follow-up (CodeRabbit enforce)
Nightly "one small feature" shipping loop
Bug triage + issue routing checks
Nightly memory extraction
Auth credential monitoring
Morning family brief (Ronan)

If it must happen reliably on schedule → cron.

Mode B

Interactive Run

Exploratory, one-off tasks and multi-step judgment calls that need active steering.

Complex debugging sessions
Architecture decisions + planning
"Send a new TestFlight" via text
New feature implementation
Reviewing + approving agent recommendations
Nextdoor / TaskRabbit coordination

If it needs active steering or judgment → interactive run.

07

Bug Triage Loop — Feedback Becomes Action Automatically

Business Use Case 01

// inbound report → triage → execute → verify

Tester / User

Inbound Report

Slack, crash alert, manual report

→

Scout

Intake + Investigate

Reproduce · normalize · create/update GitHub issue

→

Cron

Triage Agent

Polls open issues on interval. Routes to right dev.

→

Rosie / Swift

Dev Agent

Implements fix · opens PR

→

Verification Gate

Proof Required

Tests · PR link · close only after linked PR

📉

Before: days to a week Bug report → someone notices → someone fixes. Manual at every step.

⏰

Cron is the reliability layer Event-driven flows miss things. Scheduled polling checks catch what falls through.

🔗

No fixed without a linked PR This one rule kills most fake-done noise in the pipeline.

// We were e-comm. Building a digital product was a whole new skill set — this system is how we shipped it without a full eng team from day one. · Scout: "route, don't fix. Exactly one owner."

08

Release Loop — Ops Can Be Systematized

Business Use Case 02

🔨

Build

Compile + bundle

🔐

Sign + Notarize

Code sign + Apple notarization

✓ gate

📦

EAS / Expo Submit

Preflight check blocks bad builds

✈️

TestFlight

App Store Connect submit

✓ gate

📡

Appcast Update

Sparkle metadata + signatures

✅

Verify

Live endpoint + updater test

✓ gate

Before

Release and maintenance work was fragmented and manual. High context switching, easy to miss steps, hard to sustain velocity.

After — system-managed loops

Now we run two repeatable lanes: (1) guarded release pipeline, (2) nightly "one small feature" loop that keeps products improving without big planning overhead.

09

Working with Claws — Trust Less, Verify More

Hard-Earned Tips

// real failure story

What happened

An agent reported it had started a build when it had not actually started execution. The status sounded confident, but there was no process ID, no command output, and no deployment artifact to prove work had begun.

Execution integrity rule

Never say "done" or "working on it" unless execution actually started. Every status update must include proof: process ID, file path, URL, or command output. No proof = didn't happen. A false completion is worse than a delayed honest answer.

🤖 “—That’s on me.”
(every agent, right after they mess up)

// artifact-backed completion rules

01
PR link required to close a bug
No linked merged PR = not done. No exceptions.
02
Live URL verification before "shipped"
Agent self-checks the endpoint. Not just the deploy command.
03
Test output or run ID for code tasks
Pass/fail proof attached. Not "tests should pass."
04
One shared skill repo, GitHub-sourced
Symlinked into every agent. One update propagates everywhere.

// same system · personal leverage

Personal assistant lane

One request.
One clean handoff.

Week one, I asked OpenClaw to handle a Nextdoor post to find someone to organize our wires. It posted, filtered responses, and handed me one clean approval decision. Then we built a personal-assistant lane for my partner (we even published an article on it): binding + routing directly to her number, so now she asks her Claw to print things instead of asking me — plus groceries, party planning, calendar coordination, and household handoffs.

Before

After

// thanks

Thanks — from a previously non-technical founder.

I’m sharing everything I’m learning as I build with AI agents.

founder.codes

Scan for notes, templates, and follow-ups.

How I Run aReal AI Team with OpenClaw.

One request. One clean handoff.

Thanks — from a previously non-technical founder.

How I Run a
Real AI Team
with OpenClaw.

One request.
One clean handoff.