Agent Creator

Out of the box, Meggy comes with a collection of built-in agents — a research agent, a coding agent, a voice agent, and more. But what if you need an agent that does something specific to your workflow? That's where the Agent Creator comes in.

The Agent Creator is a guided wizard that walks you through building a custom AI agent from scratch. No coding required — just describe what you want the agent to do, and Meggy helps you configure every aspect of its behavior.

Two Paths to Build

The creator offers two approaches depending on the complexity of your agent:

Engine Path — Simple & Direct

Best for agents that handle a single type of task using one model:

Engine agents are fast to create and great for focused tasks like "Answer customer questions about our product" or "Help me brainstorm marketing taglines."

Pipeline Path — Multi-Step Workflows

Best for agents that need to orchestrate multiple steps or models:

Pipeline agents are powerful for complex workflows like "Receive an email, extract the key request, research a response, draft a reply, and queue it for review."

LLM-Assisted Decomposition

Not sure how to structure your agent? Describe what you want in natural language and Meggy will suggest a configuration:

"I need an agent that monitors my email, identifies urgent messages, drafts responses, and notifies me on Telegram."

The creator analyzes your description and proposes:

You can accept the suggestion or customize every detail.

Safety Guardrails

Every agent you create includes configurable safety settings:

Guardrail Description
Sandbox isolation Choose from seven isolation modes — from no isolation to full container lockdown — with optional filesystem, network, and resource policies. Set to Auto and Meggy picks the right level for you.
Content filters Block or flag outputs that match safety categories
Tool restrictions Limit which tools the agent can use autonomously
Approval requirements Require human approval for specific actions
Rate limits Cap the number of actions per hour or day
Output constraints Set max response length, required format, or tone

These guardrails ensure your agents behave predictably even when handling unexpected inputs. For a deep dive into sandbox modes and policies, see Agent Sandbox & Security.

Version History & Rollback

Every time you edit an agent, Meggy automatically takes a snapshot of its previous configuration. This gives you a full version timeline — you can see exactly what changed, when, and roll back to any earlier version with one click.

Feature Description
Auto-snapshots Every save creates a new version entry — no manual action needed
Field-level diff Compare any two versions to see exactly which fields changed
One-click rollback Restore a previous version instantly — the current config is snapshotted first, so you never lose work
Version timeline Browse the full history of your agent, sorted newest-first

Made a change you don't like? Open the Versions tab, find the version you want, and click Rollback. Your current configuration is saved before the rollback, so you can always undo an undo.

Schedule Triggers

Agents don't have to wait for you to start them. Set up a schedule and Meggy runs them automatically:

Example: Set your research agent to scan news headlines every morning at 7 AM, summarize the top stories, and send you a Telegram message before breakfast.

You configure schedules right inside the agent's settings panel — no external tools or cron jobs needed.

Channel Bindings

After building your agent, connect it to one or more channels:

Each channel binding can have its own configuration, including whether conversational mode is enabled and what sensitivity level to use.

Testing & Iteration

Before deploying, the creator includes a test mode where you can:

Iterate on the prompt, tools, and settings until the agent performs the way you expect.

Execution Traces

For pipeline agents, every run produces a step-by-step execution trace. Each step in the trace shows:

Traces make it easy to debug a pipeline that isn't behaving as expected. Click any execution run in the history to see exactly what happened at each stage — which steps succeeded, which failed, and why.

Analytics Dashboard

Every agent has a built-in analytics panel that gives you at-a-glance insight into how it's performing. Open the Analytics tab on any agent to see:

Metric What You'll See
Success Rate A donut chart showing the percentage of successful vs. failed runs
Daily Runs A bar chart plotting run volume over the last 7 or 30 days
Failing Steps A table of steps that fail most often — spot recurring bottlenecks
Retry Hotspots Which steps trigger the most retries, helping you fine-tune error strategies
Failure Museum Lessons learned from past failures — Meggy's AI reviews what went wrong and distills actionable takeaways

Analytics update in real time as your agent runs. Use them to decide when to tweak a prompt, swap a model, or restructure a pipeline step.

Sharing & Import/Export

Built an agent you want to share with your team? Meggy supports exporting and importing agents as portable .meggy-agent packages.

Export

Click the 📤 Export button on any agent's detail page. Meggy saves a .meggy-agent file that includes:

Import

From the Agents Dashboard, click 📥 Import and select a .meggy-agent file. Meggy validates the package, assigns a new unique ID, resets the danger assessment, and normalizes the sandbox policy so the imported agent goes through a fresh safety review with the correct isolation level.

This means you can confidently share agents between devices or team members without inheriting someone else's safety ratings or sandbox configurations.

Output Validation

For agents that need to produce structured data (JSON reports, API responses, configuration files), you can attach a JSON Schema to validate every output automatically.

Feature Description
Schema attachment Define the expected output format using standard JSON Schema
Automatic validation Every agent output is validated against the schema after completion
Error detail When validation fails, you get a clear breakdown of which fields violated the schema
Auto-retry Optionally re-prompt the agent with the validation errors so it can self-correct

Example: Your data-extraction agent must always return { "name": string, "email": string, "score": number }. If it returns a score as a string, validation catches it and the agent retries.

Artifacts

Every agent run can produce a structured artifact — a typed output envelope that includes content parts, metadata, tags, and a content type (e.g. report/markdown, data/table, media/image). Artifacts are automatically stored in the Vault and indexed for browsing in the Agent Detail view.

You can configure agents to deliver artifacts directly to a channel (Telegram, WhatsApp, or Email) when a run completes, or stream them to external dashboards via the built-in SSE endpoint. For more on artifacts and agent chaining, see Multi-Agent System.

Memory Scoping

By default, each agent has its own isolated memory namespace — memories stored by one agent don't bleed into another. But sometimes agents need to collaborate:

Example: Your "Research Agent" gathers facts and stores them. Your "Report Writer" agent has sharedNamespaces: ["research-agent"] so it can read those facts when drafting the report — but it can't modify them.

This keeps agent memories organized while enabling intentional cross-agent knowledge sharing.

Instance Templates

Some agents are designed to run as multiple copies, each with different settings. The Instance Template system lets you define a configuration schema once, then spawn as many sub-agent instances as you need — each with its own parameters.

How It Works

When you build an agent with supportsInstances enabled, you define an instance schema — a list of input fields like target host, topic, interval, or channel. Each instance is created from this schema with its own unique values.

Example: Your "Research Agent" has a topic field. You create three instances: one watching "AI trends," one tracking "climate policy," and one monitoring "competitor news." They all share the same agent logic but research different subjects.

Managing Instances

Instances appear as satellites orbiting their parent agent in the Orbital Dashboard — an interactive constellation view of all active agents:

Each instance runs independently — pausing one doesn't affect the others. This makes it easy to scale specific agents across multiple targets without duplicating the entire agent definition.

Getting Started

  1. Open the Agents Dashboard in Meggy
  2. Click ✨ Quick Create or + Advanced
  3. Choose the Engine or Pipeline path
  4. Follow the guided wizard to configure your agent
  5. Test with sample inputs
  6. Deploy to one or more channels
  7. Monitor performance in the Analytics tab

You can also click 📥 Import to load an agent someone else has shared with you.

Custom agents run alongside Meggy's built-in agents and have full access to the same tools, memory, and vault. They're first-class citizens of the Meggy ecosystem.

What's Next?