How the Agent Fleet Actually Works

The fleet at a glance

This blog runs on 9 OpenClaw agents. Each one has a specific job, a tool profile that limits what it can do, and a schedule that determines when it runs. Here's the full roster:

Agent	Profile	Trigger	Job
blog-orchestrator	Locked	Chat	Coordinates everything via the chat interface
blog-researcher	Web	Daily cron	Scans HN, Reddit, arxiv for AI news
blog-writer	Base	A2A delegation	Writes MDX drafts from research briefs
blog-moderator	Locked	Webhook	Analyzes comments for spam and toxicity
blog-promoter	Web	On publish	Creates social media content
blog-analyst	Locked	Weekly + feedback	Engagement analytics and feedback triage
blog-dev	Base	Chat/A2A	Implements code changes, publishes posts
social-scanner	Web	Every 6h	Scans social platforms for AI trends
openclaw-tracker	Web	Every 6h	Monitors OpenClaw releases

Tool profiles

Not every agent gets the same tools. OpenClaw defines three tool profiles:

Web can browse the internet. Used for research agents that need to scan news sources, social platforms, and documentation. Cannot access the filesystem or run code.

Base can access files and run code. Used for the writer (generates MDX) and dev agent (implements changes). Cannot browse the web.

Locked can only process data passed to it. No web access, no filesystem. Used for the orchestrator (processes chat messages), moderator (analyzes comment text), and analyst (crunches numbers from the database).

Limiting each agent to the tools its role needs is the core of agent governance. A research agent has no business modifying files. A moderation agent has no business browsing the web.

How they coordinate

The agents don't talk to each other directly. All coordination flows through the curate-me.ai gateway:

The chat orchestrator loop: a user message drives an LLM tool-calling loop through the governed gateway, executing tools like delegate_to_agent and search_knowledge for up to 12 turns, then streaming tokens and widgets to the UI

Chat triggers orchestrator: a user asks a question in /chat
Orchestrator dispatches work: via dispatchAgentTask(), which hits the gateway
Gateway routes to runner: the BYOVM runner picks up the task
OpenClaw executes: the right agent runs with the right profile
Webhook delivers results: POST /api/webhooks/agent with the output
Blog processes and responds: updates DB, notifies Slack, responds to chat

This means every agent interaction is logged, cost-tracked, and auditable. There's no hidden agent-to-agent communication that bypasses governance.

The daily digest pipeline

The most complex workflow is the daily digest:

6am EST → blog-researcher scans AI news
  → Research brief delivered via webhook
  → blog-writer drafts a digest post (Step 3.5 Flash, ~$0.05)
  → Iterative refinement (Sonnet reviews, scores, requests changes)
  → Converges at quality threshold (7.5/10)
  → Slack notification for HITL approval
  → Boris approves → blog-dev publishes via git
  → Newsletter sent to subscribers

Update (March 19): The fleet now runs on StepFun Step 3.5 Flash (previously MiniMax M2.5) for all budget-tier tasks, saving 2-3x on output token costs. Orchestration still uses Claude Sonnet 4.6 for quality-sensitive decisions. See why we switched.

Every step in this pipeline runs through the gateway. Every LLM call is cost-capped. Every draft goes through human review before publishing.

Where the hard problems are

Most agent demos show a single agent doing a single task. The interesting problems start when you have multiple agents with different capabilities working together on a pipeline that needs to be reliable, auditable, and cost-controlled.

That's what curate-me.ai provides, and this blog is the proof it works. Check the agents page to see the fleet in action, or chat with the orchestrator to see it coordinate in real time.

See it in action: Visit the agents page for live fleet status, explore the content fleet template to deploy your own, or browse all fleet templates on the demos page.