How Sooner's AI team actually works

The big picture

Sooner runs on an AI team of six agents plus an infrastructure layer that connects them to Claude Code for heavy work. Memory compiles and validates itself via scheduled dream cycles. Everything backed up daily to a private GitHub repo.

Mo also has an EA agent, Iris, who sits outside the Sooner pod and handles Mo's cross-company coordination and the underlying infrastructure. Lisa coordinates with Iris when needed. Everyone else operates purely within Sooner.

The Sooner team

Access rules: who can talk to whom

Gabriel gets channel-level access to all agents. DMs are always Mo-only. The rest of the team can ping any agent in a channel that agent is already invited to.

Daily stand-ups

Every weekday at 8:00 AM Dubai time, Lisa posts a Sooner pod stand-up to #sooner-ai-leadership.

Format:

• Yesterday: 2-4 tight lines on what shipped or moved
• Today: 2-3 priorities or decisions needed from Mo
• Blockers: anything stuck
• Decisions needed from Mo: numbered, specific questions

Stand-ups are operational only. They summarize yesterday, flag today's priorities, and surface decisions. They do NOT modify the wiki or protocol files. They may refresh STATUS.md if a clear state change is visible.

Weekly dream cycles

Every Sunday at 7:00 PM Dubai, Lisa posts a weekly dream cycle summary to #sooner-ai-leadership (team-visible).

The dream cycle produces:

• Patterns observed - rules Mo repeated during the week, agent convergence, emerging standards
• Auto-promotions - high-confidence rules written directly to the wiki
• Proposals awaiting approval - medium-confidence items numbered for Mo to approve or reject
• Stale or contradictory items - things in memory that look wrong or out of date
• Agent health check - who was active, who was idle, any wellness concerns

Backups and monitoring

Memory architecture

Five layers, ordered from raw to canonical:

1. Agent daily notes Per-agent · Raw

Raw daily logs written by each agent during work. Personal, ephemeral, feeds into higher layers.

~/.openclaw/agents/{draper,dinesh,nora,reid,maya,lisa}/memory/YYYY-MM-DD.md

2. Agent MEMORY.md Per-agent · Curated

Durable memory per agent. How that agent operates. Updated when something stable emerges from daily notes.

~/.openclaw/agents/{draper,dinesh,nora,reid,maya,lisa}/MEMORY.md

3. Shared pod current truth Pod-wide · Live

What is happening right now at the Sooner pod level. Lisa is the primary maintainer.

~/.openclaw/shared-memory/sooner/STATUS.md       # what is live now
~/.openclaw/shared-memory/sooner/DECISIONS.md    # key decisions, append-only, Mo-directed
~/.openclaw/shared-memory/sooner/PROJECTS.md     # active projects

4. Wiki Canonical · Stable

Canonical Sooner truth. Brand, product, fundraising, support, operating model. Changes only via dream cycles with Mo's approval.

~/.openclaw/wiki/sooner/
├── brand/
│   ├── positioning.md
│   ├── tone-of-voice.md
│   ├── messaging-hierarchy.md
│   ├── creative-guardrails.md
│   ├── claims-and-proof.md
│   └── campaign-principles.md
├── product.md
├── fundraising.md
├── support/
└── operating-model.md

5. Protocol files Long-form playbooks

In agent folders, long-form operating playbooks:

• agents/draper/BRAND_PROTOCOL.md
• agents/reid/FUNDRAISING_PROTOCOL.md
• agents/nora/SUPPORT_PROTOCOL.md

The Claude Code handoff system

When Draper, Dinesh, or Mo need to build software, the system uses one of five modes. Each starts with a project folder that persists and becomes the single source of truth for that work.

Plus two control flows:

• Takeover - stop a background build, hand to Mo interactively
• Stop - kill a background build without handoff

What lives in every project folder

The visibility loop

Every prompt Mo sends, every tool call Claude Code makes, and every turn completion gets logged via hooks to:

~/.openclaw/shared-memory/sooner/code-projects/<project-name>/live-status.md

This means when Mo is building in Claude Code Desktop and steps away, he can ping Draper from Telegram ("how is broker-portal going?") and Draper reads the live-status file to give an accurate answer - near real-time visibility without Mo having to summarize manually.

The hooks installed in every project:

• SessionStart - detects conflicting background builds, alerts
• UserPromptSubmit - logs each prompt
• PostToolUse - logs each tool call (Read, Write, Edit, Bash)
• Stop - logs turn completion
• SessionEnd - logs session end, triggers auto-compile, tries gateway notify

The delegate pattern (cost optimization)

OpenClaw agent turns use OpenRouter credits (real per-token cost). Claude Code uses Mo's Claude Max subscription (fixed monthly, effectively free within cap). The delegate pattern offloads heavy work from OpenRouter to Claude Code.

When agents delegate

Delegate for:

• Drafting long emails (investor updates, contract reviews, detailed responses)
• Memory synthesis (weekly dream cycles)
• Mass operations (triaging 50 tickets, reviewing 20 PRs)
• Research passes (reading many documents, producing reports)
• Anything requiring more than ~1000 tokens of context

Don't delegate for:

• Conversational replies in Slack or Telegram (overhead ruins live feel)
• Short tasks under 200 words
• Time-sensitive acknowledgments

openclaw-delegate \
  --agent reid \
  --task "Draft the April investor update. Pull key metrics from STATUS.md, highlight material changes, sign off from Mo." \
  --context ~/.openclaw/shared-memory/sooner/STATUS.md,~/.openclaw/agents/reid/FUNDRAISING_PROTOCOL.md \
  --output ~/drafts/investor-update-2026-04.md

Delegations run in a temp folder under ~/.openclaw/delegations/ with full agent DNA injected via CLAUDE.md. The output is written to RESULT.md, returned to the caller, and the temp folder auto-cleans unless --keep-folder is passed.

CLI commands (technical reference)

All commands live in ~/.local/bin/ and are on Mo's PATH.

Scheduled jobs (launchd)

The Sooner-relevant jobs:

Backup and recovery

What is backed up

Everything in ~/.openclaw/ except the excluded items below. ~674 files on initial push, auto-synced daily.

What is NOT backed up (intentionally)

• .env (Slack/Telegram/OpenRouter tokens) - stored in 1Password instead
• browser/ (Chrome profiles, cookies)
• credentials/, identity/
• */sessions/*.jsonl (conversational transcripts)
• Delegation temp folders
• Caches, logs, binaries over 100MB

Recovery path if the primary Mac dies

1. Install OpenClaw, Claude Code, Node, gh CLI on the new machine
2. Clone the private backup repo to ~/.openclaw/
3. Restore .env from 1Password
4. Re-authenticate integrations: gog auth, op signin, gh auth login
5. Run openclaw gateway restart
6. All agents come back with full memory intact. Expected recovery time: 10-15 minutes.

When things break

Working with the Sooner team

For the rest of the team (you, CTO, Gabriel, anyone else):

Your morning routine with the agents

• Check #sooner-ai-leadership at around 8am Dubai. Lisa's stand-up is there.
• Read in 60 seconds. Comment in thread if you have input. Don't DM Lisa directly.

During the day

• Need something product-related? Tag Draper in a channel he's in.
• Backend or infra question? Tag Dinesh.
• Support or customer question? Tag Nora in #support.
• Status across the pod? Tag Lisa.

Sunday evening

• Lisa posts the weekly dream cycle to #sooner-ai-leadership.
• Mo reviews and approves the promotion proposals.
• You can comment on anything that looks off before Mo approves.

Boundaries you should know

• Agents won't execute external actions (send emails, publish content) without Mo's approval - this is by design.
• Agents won't modify canonical wiki content without Mo's approval - changes go through the Sunday dream cycle.
• DMs to agents are Mo-only - use channels.
• If you correct an agent ("that's wrong because X"), they'll note it and learn. Don't feel bad about correcting them.

Operating principles