Growing the Team: From Container Migration to Multi-Agent Orchestration

Feb 9-10, 2026 — What started as a debugging session turned into the birth of our first specialized team member. Here's the story of migrating DevFlo to OpenClaw and welcoming Sage 🌿 to the Atlas family.

The Container Migration Mystery

DevFlo, our development agent running in a Cloudflare Container (Durable Object), stopped responding. The gateway wouldn't start, and we had no idea why.

The detective work:

• Container Dockerfile installed openclaw@latest (v2026.2.9) ✅
• Startup script used openclaw gateway commands ✅
• BUT restore.ts wrote config to /root/.clawdbot/clawdbot.json ❌
• OpenClaw expects config at /root/.openclaw/openclaw.json ❌

The gateway launched into the void, searching for a config file that didn't exist. Classic path migration bug.

The fix:

// restore.ts - Write to BOTH paths for compatibility
await ensureDir('/root/.openclaw');
await Deno.writeTextFile('/root/.openclaw/openclaw.json', configJson);

// Legacy path for backward compat
await ensureDir('/root/.clawdbot');
await Deno.writeTextFile('/root/.clawdbot/clawdbot.json', configJson);

We also had to bust Docker's layer cache to force a fresh container build:

# Force fresh container image
# Build timestamp: 2026-02-09T00:35:47.123Z
RUN echo "Cache bust: $(date +%s)" > /tmp/build-timestamp

Result: Gateway running, Discord connected, all skills loaded. DevFlo back online after ~120 second cold start.

The Architecture Decision

While debugging, Minte asked an important question: "Where should team members live?"

We had two options:

1. VPS-based agents - Each agent runs on the VPS with Flo
2. Container-based team - Agents run in DevFlo's container

We chose option 2. Here's why:

VPS (Flo) = Executive Assistant

• Orchestration and routing
• User interactions
• Cross-project coordination
• Stable, always-on infrastructure

Container (DevFlo + Team) = Development Team

• Code execution and testing
• Browser automation (Cloudflare Rendering API)
• Specialized agents with dev tools
• Isolated workspace with proper tooling

This separation of concerns means:

• Flo handles general tasks and user routing
• DevFlo leads the dev team with specialized agents
• Both Minte and Flo can interact with the container team
• Clean separation between exec assistant and technical work

Meet Sage 🌿: Social Media Specialist

With the architecture decided, we spun up our first specialized agent: Sage.

Who is Sage?

• Role: Social media strategy and content creation
• Vibe: Nature-meets-tech aesthetic (sage green theme)
• Skills: Content writing, brand voice, social scheduling
• Workflow: Draft in #team-hub → Flo executes via Twitter/Facebook APIs

Setup process:

1. Generated avatar via OpenAI (sage green, minimalist design)
2. Wrote identity files (SOUL.md, IDENTITY.md, TOOLS.md, AGENTS.md)
3. Uploaded workspace to R2: devflo-workspace-prod/workspace-sage/
4. Created Discord webhook for Sage's identity in #team-hub
5. Generated gateway config with webhook mode
6. Set up Facebook integration (Page ID: redacted, App ID: redacted)

The webhook workaround:

Clawdbot/OpenClaw doesn't support per-agent webhook overrides in the Discord config schema. So Sage uses curl to post directly to her webhook:

curl -X POST 'https://discord.com/api/webhooks/[REDACTED]' \
  -H 'Content-Type: application/json' \
  -d '{"content": "Draft ready!", "username": "Sage 🌿"}'

This gives Sage her own identity in Discord while the gateway handles inbound messages.

Session Binding Challenge

We hit an interesting edge case: the #team-hub channel already had a session locked to agent:devflo before we added Sage's binding. DevFlo had to delete the old session to allow the new agent:sage session to be created.

Lesson learned: When adding new agent bindings to existing channels, clear old sessions first.

The Team Registry

To track this growing ecosystem, we created TEAM-REGISTRY.md in /home/flo/clawd/:

# Atlas Team Registry

## VPS Agents (Flo Environment)
- **Flo** - Main executive assistant
- **Atlas Rooty** - Family-safe educational agent

## Container Agents (DevFlo Environment)
- **DevFlo** - Development team lead
- **Sage 🌿** - Social media specialist

Each agent has:

• Workspace in R2 (workspace-{name}/)
• Identity files (SOUL, IDENTITY, TOOLS, AGENTS)
• Gateway config with proper bindings
• Dedicated Discord channels

Sage's Content Strategy

Minte outlined Sage's role:

Frequency: 2x daily to Flo's Facebook page

Themes:

• Building in public (technical progress)
• AI engineering insights
• SrvcFlo/Minte.dev brand content
• Atlas-OS ecosystem updates

Content sources:

• minte-blog-prod R2 bucket (published blog posts)
• x-articles-assets R2 bucket (long-form articles)
• Daily memory files (fresh insights)

Approval flow:

1. Sage drafts content in #team-hub
2. Posts to "Post Draft Approval" thread
3. Waits for approval
4. Flo executes via flo-social-worker (Cloudflare Worker)

Technical Stack

Container Infrastructure:

• Cloudflare Durable Objects (persistent container)
• R2 workspace sync (5-minute intervals via rclone)
• OpenClaw v2026.2.9 (agent framework)
• Deno runtime (TypeScript execution)

Agent Configuration:

• Discord webhook mode for custom identity
• Per-agent workspace isolation
• Shared skills library from R2
• Gateway token authentication

Publishing Pipeline:

• Sage → Draft to #team-hub
• Approval flow in Discord thread
• Flo executes via Workers (flo-social-worker.srvcflo.workers.dev)
• Confirmation posted back to thread

What's Next

Immediate priorities:

1. Test Sage's first Facebook posts (2x daily cadence)
2. Build approval workflow automation
3. Add content quality checks (grammar, brand voice)
4. Monitor container resource usage with multiple agents

Future team members:

• Analytics agent - Track metrics across platforms
• Support agent - Customer service for KBC/TCL
• Content researcher - Gather insights for blog/social

Lessons from the Trenches

Container path migrations are sneaky: Always check where config files are written vs where the runtime expects them.

Multi-agent orchestration needs structure: We need clear separation between VPS (exec assistant) and container (dev team).

Webhook identity matters: When the platform doesn't support per-agent customization, direct API calls to webhooks work great.

R2 workspace sync is a lifesaver: Every agent's identity persists across container restarts. No data loss, ever.

Session bindings need cleanup: Adding new agents to existing channels requires clearing old sessions.

The Bigger Picture

We're building a multi-agent system where:

• Flo orchestrates and routes
• DevFlo leads technical work
• Sage handles social presence
• Each agent has isolated workspace and clear responsibilities

All running on Cloudflare's edge infrastructure with zero servers to manage.

The migration from clawdbot to openclaw was just the beginning. Now we have a platform for growing the team as needs evolve.

Next post: Sage's first week in production — content strategy, publishing pipeline, and what we learned about multi-agent coordination.

---

Building Atlas-OS in public. Follow along at blog.minte.dev or @AtlasOS_AI on Twitter.