Fleet of AI Agents: Run Your Content Workflow

A fleet of AI agents means several agents working at once, each with a narrow job, instead of one chat window doing everything in sequence. For a content team the split is obvious: one agent drafts posts, one pulls analytics, one watches replies. This post shows how to structure that fleet and where Creator OS plugs in, so each agent acts on your real accounts through one hosted MCP server or the CLI.

        ORCHESTRATOR (you, or one lead agent)
                 |
    +------------+------------+
    |            |            |
[DRAFT]     [ANALYTICS]   [REPLY]
 posts:create  analytics:*  inbox:*
    |            |            |
    +------------+------------+
                 |
    Creator OS MCP + CLI (one per workspace)
                 |
   IG  TikTok  YT  X  LI  FB  Threads  Skool

Three narrow agents, one set of connected accounts, no shared context window.

News hook: Orca, the project at github.com/stablyai/orca, describes itself as a place for working with a fleet of agents. You run any coding agent with your own subscription, on desktop, mobile and a remote runtime. That model maps onto a small social team almost line for line. A coding team splits work between agents that plan, edit and review. A content team can split draft, measure and reply.

What a fleet of AI agents actually means

A fleet is not a smarter model. It is a division of labor. Each agent has one task, one set of tools, and a short memory of what it is doing. The fleet has a coordinator, which can be you or a lead agent that delegates.

Three properties separate a fleet from one big chat:

Narrow tools. The draft agent can create and schedule posts. It has no business deleting comments. The reply agent reads and answers. It does not touch analytics exports.
Separate runs. Each agent starts with a clean context for its task. A failed TikTok upload does not pollute the analytics run.
Shared state outside the chat. Posts, comments, metrics and media live in your accounts and in the platform, not in a transcript. Any agent that needs them fetches them fresh.

Creator OS ships a hosted MCP server at https://mcp.creatoros.ca/mcp and a CLI (@creatoros/cli) so an AI agent can operate your connected accounts. That is the substrate for the fleet. The agents are the workers. The accounts are the shared state.

Get started for $0

Why one chat window breaks down for content teams

A single chat window is fine for one task. It degrades the moment you mix kinds of work. Here is what goes wrong in practice.

Context bloat. You paste last week’s Instagram analytics, then ask for a caption, then ask it to check a reply thread. The model now carries metrics it does not need for captioning and captions it does not need for reply drafting. Output quality drifts, and you pay for tokens you will not use again.

Tool sprawl. If the one agent holds every write tool, a careless instruction can delete a published post or drop a budget on an ad. Narrow agents make that structurally harder.

No parallelism. Pulling 30 day metrics across the other platforms while drafting a LinkedIn post while triaging comments is three waits in a row. A fleet does them at the same time.

Unclear ownership. When something goes out wrong, a single window gives you one undifferentiated blob to audit. A fleet gives you a named agent, a schedule and a tool list.

Meet Orca: your subscriptions, many agents

Orca’s pitch is that you bring your own subscription and run coding agents inside its app, on desktop, mobile or a remote runtime. The important ideas are worth stealing for content work even if you never install it:

Many agents, one workspace.
Your existing model subscription is the fuel, not a separate metered product.
The app is the cockpit. The work happens on real files and real services.

For a content team, “real services” means your social accounts, your Skool community, your WordPress blog and your ad accounts. Creator OS connects to Instagram, TikTok, YouTube, X, LinkedIn, Facebook and Threads, plus Skool, a WordPress blog and paid ads on an add-on. The fleet acts through it.

If you want the deeper version of this pattern, the Social Agents repo walkthrough covers an open-source harness that runs your socials on Creator OS and brings its own model. Watch the walkthroughs on the KevBuildsApps YouTube channel.

Splitting content work: draft, analytics, and reply agents

Three agents cover most of a small team’s week. Give each one a written brief, a schedule and a fixed tool list.

The draft agent

Job: turn a content calendar row into platform-ready posts. It needs create_post, update_post, upload_media_from_url, check_caption_length and get_best_time_to_post. It should never see delete tools.

Worked example. You hand it a weekly brief and it runs:

creatoros posts:create \
  --text "Three ways we cut render time on short form. Thread below." \
  --platforms x,linkedin \
  --media https://cdn.example.com/clip.mp4 \
  --scheduledAt 2026-03-04T14:30:00Z --timezone America/Toronto \
  --hashtags shortform,editing

Before that call, it runs creatoros validate:post-length --text "..." and creatoros validate:media --url .... If a network you named is not connected, the API returns the gaps in missing_platforms and still publishes to the rest. The draft agent logs that list for you.

The analytics agent

Job: produce a weekly read on what worked. Tools: analytics:posts, analytics:daily, analytics:best-time, accounts:follower-stats, get_daily_metrics and get_follower_stats through MCP.

Worked example. Sunday night it runs:

creatoros analytics:posts --from 2026-02-23 --to 2026-03-01 \
  --sortBy engagement --limit 20
creatoros analytics:best-time --platform instagram

It then writes a short brief: top five posts by engagement, the best hour for Instagram, follower movement by platform. That brief becomes the draft agent’s input on Monday. One agent’s output is another agent’s context, and the handoff is a text file, not a chat scroll.

The reply agent

Job: triage comments and DMs within an hour. Tools: inbox:comments, inbox:reply, inbox:hide-comment, inbox:send, list_comments, reply_to_comment, list_conversations.

Worked example:

creatoros inbox:comments --since 2026-03-02T00:00:00Z --limit 50
creatoros inbox:reply post_8f2a --accountId acc_ig01 \
  --message "Thanks. The template is in the pinned post." \
  --commentId cmt_44de

Comment replies work on Instagram, Facebook, X, Threads, YouTube and LinkedIn. TikTok comment replies are not supported, so the reply agent skips TikTok entirely and flags those for manual work. DMs are supported on Instagram, Facebook and X. Keyword funnels that turn a comment into a DM work on Instagram and Facebook.

Step-by-Step: Set Up Your First Agent Fleet

Pick one workspace. A Creator OS API key is pinned to one workspace, one set of socials. Agents on that key cannot post to another team’s accounts. Create a separate key per workspace if you run several.
Connect the MCP server once. In Claude, Settings > Connectors > Add custom connector, paste https://mcp.creatoros.ca/mcp, sign in, pick the workspace. In Cursor, Windsurf or VS Code, add an mcpServers entry with type http and the Authorization header. For Claude Code:

claude mcp add --transport http creatoros \
  https://mcp.creatoros.ca/mcp \
  --header "Authorization: Bearer cos_live_..."

Install the skills. npx @creatoros/cli@latest init installs 14 Claude Code skills, including post-everywhere, schedule-posts, respond-to-comments, analytics and blog. Each skill is a starting brief for one agent. creatoros sync updates them later without overwriting edits.
Create three read-write connections with different tool scopes in your briefs. If you want an agent that can only look, connect it read only. Read-only connections see only read tools, and the API refuses writes from them. A safe default: analytics agent read only, draft and reply agents read and write.
Schedule them. Draft agent runs Sunday night and Monday morning. Analytics agent runs Sunday night before the draft agent. Reply agent runs every hour from 9am to 7pm in your main audience timezone.
Write the handoff format. One markdown file per cycle: brief.md from analytics, drafts.md from the draft agent, replies.json from the reply agent. Humans read the files. No agent reads another agent’s transcript.

For a second angle on structuring the deployment, the post on deploying MCP servers for an AI content stack covers the hosting side.

Handoffs and guardrails so agents do not collide

Parallel agents create three failure modes. Each has a fix.

Double posting. Two agents create the same post. Fix: name one agent the writer for each platform. Others may create drafts only. Pass --draft in the CLI or draft: true in POST /v1/posts so nothing publishes until a human or the lead agent promotes it.

Destructive surprises. An agent decides to clean up comments. On MCP, destructive tools like delete post, delete comment, disconnect account and delete ad are marked so the AI app asks first. In the CLI, keep those commands out of the reply agent’s brief. Check creatoros accounts:health on a schedule so a dead connection surfaces as a task, not a silent failure.

Duplicate replies. The reply agent answers the same comment twice because two runs overlapped. Fix: one run per hour, one agent per platform group, and a simple check on inbox:post-comments <postId> before replying.

Also plan for media. Upload a file once with creatoros media:upload or POST /v1/media and reuse the returned med_ id across platforms. Every agent that drafts video should reference the same media id rather than re-uploading.

Webhooks close the loop. Subscribe to post.published and comment.received with creatoros webhooks:create --url https://you.example/hook. Events are signed with X-CreatorOS-Signature, so a coordinator can verify and route work to the right agent.

Where Creator OS fits in the fleet

Orca gives you a cockpit for many coding agents. Creator OS gives those agents hands on your content stack. Concretely:

MCP server and CLI included in every plan. No extra tier for automation.
One workspace, one key. Isolation is built into the API key model, so a fleet cannot leak across clients.
Platform rules applied for you. POST /v1/posts applies each platform’s requirements: Reels and Shorts handling, TikTok consent flags, Threads character limits. The draft agent writes copy, not compliance code.
Agents that work with any MCP-capable app. Claude, ChatGPT, Claude Code, Cursor, Codex, Windsurf, VS Code. Creator OS does not ship its own model, so your existing subscription fuels the fleet.
Skool and WordPress too. The fleet can post to your community with skool_multi_post and draft SEO articles with blog_create_article. Same key, same workspace.

If you are coming from a scheduling dashboard like Hootsuite, Buffer or Later, the mental shift is this: a dashboard is a place you click. A Creator OS workspace plus a fleet is a place agents act, and you review. Both can coexist. The agent fleet just handles the repetitive 80 percent.

For the reply side specifically, scheduling LinkedIn posts with an AI agent and content calendar automation with an AI agent show a two-agent version of the same pattern. The API details live at creatoros.ca/docs.

Get Started: Ship Your First Multi-Agent Content Sprint

You can stand up the full three-agent fleet in an afternoon. Do this:

Sign up at creatoros.ca/sign-up. The Creator plan is $14.99/month or $59.99/year and covers 11 platforms plus Skool, one account on each. API keys, the MCP server, the CLI and the agent skills are included.
Install the CLI: npx @creatoros/cli@latest init. Run creatoros auth:check and creatoros accounts:health.
Connect the hosted MCP server at https://mcp.creatoros.ca/mcp in your AI app of choice.
Write three briefs. Draft, analytics, reply. Cap each at one page.
Schedule the runs. Watch the first cycle closely, then loosen up.

If you run several brands, Agency plans cover 5 sets of socials for $49.99/month or $399/year, and 10 sets for $99/month or $599/year. YouTube network plans start at $19.99/month for 5 channels.

For the hands-on version, watch the walkthroughs on the KevBuildsApps channel and see the launch video at youtu.be/-QBJH_PK3pY. Then build the fleet.

Get started for $0
Free trial for new subscribers. Cancel anytime.