Creating Agents via MCP: Draft, Review, and Activate

Overview

When you create a prompt-based agent using the GoMarble MCP manage_agents tool, the agent is saved in draft mode by default. This means the agent is persisted in your account but is not running until you explicitly activate it. This gives you a chance to review the agent's configuration before it goes live.

Note: Report-based agents are not affected by this change — they are always created as live.

What you can do with manage_agents

You don't need to call the tool by name. Just ask your assistant, for example "Create a weekly agent that emails me my Meta ROAS every Monday". With manage_agents the assistant can:

• Create a report agent (a saved GoMarble report sent on a schedule) or a prompt agent (a question GoMarble answers on a schedule). Schedules can be daily, weekly or monthly.

• List your agents, including agents teammates have shared with you.

• Update an agent's prompt, schedule, delivery or settings.

• Pause and resume an agent.

• Run an agent right now.

• Delete an agent.

The Draft → Review → Activate Lifecycle

1. Create the agent via prompt. When you ask the MCP to create a new prompt-based agent, it is saved as a draft by default. You will see it appear in your account in a draft state, not yet running.

2. Review the draft. Check the agent's configuration — goals, schedule, governance settings, and any other details — before committing to activation. In GoMarble chat, each new agent appears as a card with Review & edit. From another AI client, open the agent in GoMarble to review it.

3. Activate the agent. When you are satisfied with the configuration, explicitly activate the agent: click Take agents live on the card in GoMarble chat, or ask your assistant to take the agent live. Activation routes through GoMarble's activation engine, including any required entitlement checks, and the agent will begin running from that point. If your plan doesn't include agents, activation tells you to start a trial or upgrade.

A draft can't be run with "run now" until it has been activated.

Going Live Immediately

If you want an agent to go live at the moment of creation — skipping the draft step — you can explicitly request this when instructing the MCP tool. Only use this option when you are confident in the configuration upfront.

Choosing where results go

Every agent delivers by email, Slack, or both. The assistant will ask you which one you want.

• Email works for everyone. Give the email addresses that should receive the results.

• Slack needs a connected Slack workspace. Name the channel to post in, and make sure the GoMarble bot is a member (invite it to private channels first). See How to Connect Your Slack Workspace to GoMarble.

Each delivery method needs its destination. An agent set to email with no addresses, or Slack with no channel, is rejected so it doesn't run and deliver nothing.

Prompt agents can deliver every time, or only when a condition you set is met.

Agents that make changes

By default a prompt agent only reads and reports. You can give it write access so it can change campaigns:

• A write agent must include at least one Meta Ads or Google Ads account.

• Ask for approval (the default): the agent proposes changes and waits for you to approve them.

• Apply automatically: the agent applies its changes without waiting.

• Guardrails limit what a write agent may change. You can allow only certain kinds of changes (for example only pausing) and cap budget changes at a percentage. Guardrails are checked on every change the agent makes. They need a plan with agent governance. Without it, the agent is still created and you'll see a warning. See Setting Agent Limits and Governance Policies.

Learning from past runs

Prompt agents can learn from their own past runs. The setting is auto by default, and you can turn it on or off. See Using Agent Learning to Help Your Agent Remember Past Runs.

Fleet Grouping

When creating multiple related agents as part of an onboarding setup, you can group them into a fleet. A fleet bundles related agents together so they can be reviewed and activated in a single step, rather than one by one.

To use fleet grouping, provide a fleet ID and a role for each agent when creating them through the MCP tool. Once all agents in the fleet are ready, you can activate the entire fleet at once.

Shared agents

Agents a teammate shared with you appear in your list too. What you can do depends on how it was shared:

• Viewer: you can see the agent but not change, pause, resume, run or delete it.

• Editor: you can change its settings, pause, resume and run it.

• Only the owner can delete the agent or change its permissions and guardrails.

Key Points to Remember

• Prompt-based agents default to draft mode — they will not run until activated.

• Report-based agents are always created live and are unaffected by this workflow.

• You can review an agent's full configuration before activating it.

• Fleet grouping lets you activate a set of related agents together in one step.

• Account scope: when creating an agent, you can set accountScopeStrict to true to lock the agent to the ad accounts it was configured for — denying reads and writes on any other account. This defaults to false. You will be asked to confirm this choice during agent creation, as it is a safety decision that cannot be silently changed by a later update that omits the field. It's recommended for agents that apply changes automatically, or that work on a single brand.

• Every agent needs a delivery method and a destination: email addresses, a Slack channel, or both.

• Agents that make changes must include at least one Meta Ads or Google Ads account.