Pipelines

A pipeline defines how agents collaborate on a task. It specifies the order of operations, what data flows between agents, and where humans can intervene.

Example: Support Resolution Pipeline

This is the pipeline we use for our own demo. A customer reports a content error, and four agents collaborate to fix it.

Email arrives
  ↓
Support Agent (triage + dispatch)
  ↓ posts brief to #research
Research Agent (verify the claim)
  ↓ posts findings to #research
  ↓ [timed review: 15 min]
Content Agent (fix the content)
  ↓ commits to GitHub → Vercel deploys
  ↓ posts to #ops
QA Agent (verify the deployment)
  ↓ screenshots + confirmation
  ↓ posts to #ops
Support Agent (reply to customer)
  ↓ sends email with fix summary
Done

Pipeline Components

Trigger

What starts the pipeline. Can be:

Email — new message matching a filter
Telegram message — direct message or topic post
Scheduled — cron-based (daily report, weekly audit)
Webhook — external system fires an event
Manual — you tell an agent to start

Steps

Each step defines:

Agent — which agent handles this step
Input — what data it receives from the previous step
Action — what it should do
Output — what it produces for the next step
Review gate — none, timed, or required (see Human-on-the-Loop)

Error Handling

What happens when a step fails:

Retry — try again (with backoff)
Skip — move to the next step
Escalate — notify a human and pause
Abort — stop the pipeline and notify

Building Your First Pipeline

1. Define the goal

What should happen end-to-end? Write it out in plain language first.

“When a customer emails about a pricing error, verify the error exists, fix the content, confirm the fix deployed correctly, and reply to the customer.”

2. Map agents to steps

Which agent is best suited for each part?

Step	Agent	Why
Receive + triage email	Support	It monitors the inbox
Verify the claim	Research	It can fetch and analyze web pages
Fix the content	Content	It has repo access and can commit
Verify deployment	QA	It can screenshot and compare
Reply to customer	Support	It owns the customer relationship

3. Define handoffs

What data flows between steps? Keep it structured so receiving agents can parse it reliably.

4. Set review gates

Start with “timed review” on every step. As you build confidence, remove gates from low-risk steps and tighten them on high-risk ones.

5. Test with a dry run

Send a test email and watch the pipeline execute in your Telegram monitoring channel. Intervene if anything looks off.

Advanced: Automated Development Pipeline

For teams that want their agents to build and deploy web applications:

User: "Build me a task tracker app"
  ↓
Coding Agent (scaffolds the app)
  ↓ creates project, writes code
  ↓ commits to user's repo
QA Agent (reviews + tests)
  ↓ runs tests, checks for issues
  ↓ approves or requests changes
Deploy Agent (ships it)
  ↓ deploys to user's subdomain
  ↓ e.g. username.moes.ai/apps/tasks
User gets notified: "Your app is live"

This requires:

A coding agent with access to a git repo (user’s GitHub or hosted)
A QA agent that can run tests and verify output
A deployment target (Vercel, Cloudflare Pages, or managed hosting)
Larger compute for the coding agent (builds need more resources than chat)

Users can connect their own GitHub and Vercel accounts for full control over their code and deployments. Or use Moe’s managed infrastructure for a simpler setup.

See Custom Agents for how to configure coding and deployment agents.