Model Relevance Protocol
MRP Workforce
MRP agents can get complex work done with just a few simple low-level tools, a long-context model and optional docs.
Give an MRP agent a natural language task, and it'll be able to figure out how to get the job done with just a few very simple tools.
HubSpot MRP
Generate a few records to pre-populate my CRM like a live demo environment. I want Contacts, Deals, Companies, Notes, Logged Emails and Calls. Use Dummy numbers.
Google Docs MRP
Generate a Google Doc titled “AI Agent Use-Cases and Building Techniques”. Break down each AI Agent use-case and building technique in-depth.
Webflow MRP
Generate a new blog post about AI Agents and add to the "Blog Post" collection. Return the full content for me to review and approve. Once I've approved, publish it and share the link to the live page.
Notion MRP
Generate a task management system with an inline-database for projects, another for tasks. Use a relation field to link the tasks to the projects. Add 3 x example projects.
.png)
Connect your MRP Agents together into an MRP Workforce, and watch as they collaborate to get complex work done across your entire tech stack.
MRPs and MCPs
MRP stands for Model Relevance Protocol. It is an extension to, and in some cases a replacement for MCPs (Model Context Protocol).
The Model Context Protocol standardise how tools are shared and accessed. You can think of an MCP as a bundle of tools packaged together in a way that lets AI assistants use them through a shared interface. This makes it possible for agents built on one platform, to use tools developed on another.
The Model Relevance Protocol takes things a step further. Instead of packaging multiple tools into one server, an MRP is about having one agent with a couple of low-level tools that it can use to complete complex workflows. Like a generic API or code step tool that the agent can figure out how to use, instead of 30+ tools that are hyper-specific.
MRPs allow your AI assistants to complete work that isn't possible with just the tools packaged into an MCP bundle alone. Instead of writing out every step or copying detailed API calls, you can just connect your agent to the relevant MRP agent and ask for what you want in plain language. The specialist MRP agent handles the rest.
Model Relevance Protocol
What does an MRP look like at an agent level? And why might you want to use them as an extension to, or instead of MCPs? Let's dive in.
MCP - Many good tools
Your AI assistant can use any tool that an integration provider has packaged up for you. Each tool will be very reliable, and do its job well. However, it will only be able to complete work that it has tools for. If the API changes, you have to update any tool that is affected, and any tool that require strongly formatted inputs.
MRP - 1+ simple tool
Your AI Agent can figure out how to complete complex work, by dynamically filling in the input of a low-level tool, like an API step, or a code step etc. If it gets it wrong, it can self-heal and keep trying. If the API changes, you only need to update the agent API documentation.
MRP workforces, are mindblowing
It avoids overcrowding a single agent with 30+ tools, by having a single generic API tool, or code step, or LLM step etc that the agent can figure out how to use.
When the API changes, the only thing you need to change is the documentation in the agent's knowledge, instead of any of the 30+ tools that are affected by the change.
Example MRP workforce task
After cloning the multi-agent MRP workforce below and connecting our accounts to each tool, we were able to get the workforce to do the following, without building a single custom tool 🤯
Then, the workforce did all of the work below. This would have taken weeks of work to build out all of the individual tools for each integration previously. All we needed to do was connect our accounts and write a task prompt. That was it.
The MRP manager started by delegating to the research assistant to find out everything it could about Relevance AI, using two simple tools: Google Search and Extract Website Content.
Then the MRP Manager asked the HubSpot assistant to retrieve all of the records in HubSpot that are related to Relevance AI, it's founders, engagement history and key stakeholders.
The HubSpot assistant (ahem, wizard) did a very good job of this:
Next, the MRP manager asked the Brandfetch assistant to fetch all of Relevance AI's brand assets. It found logo files in different formats, brand colour codes, social profiles and more.
After that, MRP manager asked the Apollo Assistant to get all of the key stakeholders and decision makers from Relevance. It was able to retrieve a much more accurate and recent list than what was already stored in the CRM about them.
Next, the HubSpot assistant was able to update the CRM with all of the new information about the company and stakeholders that were discovered using all of these different integrations. With that, the first three requests the user made right at the start have been complete. Four more remaining.
Now, the MRP agent asks the Gamma assistant to create a slide deck for the meeting with Jacky Koh, one of the founders of Relevance AI. It asked it for 14 slides, with a mix of visuals, diagrams and non technical language.
Here are some of the slides in the 14-paged slide deck it created. Gamma is amazing. With that, the users third request for a slide deck that communicates the benefits of AI agents and cater to Relevance AI is complete.
Next, the MRP agent delegates to the Notion assistant, to create a page in Notion to prepare all of the latest details about Relevance AI ready for the meeting. The Notion Assistant failed in this case, because it required a page ID. It didn't stop working though. It carried on and finished the rest of its objectives.
When everything else was done, I sent a follow-up message to the workforce and gave it the correct page ID, which meant it was able to complete it's Notion objective too. I know for next time to give it a page ID, or ask it to pick one itself. The user also requested a spreadsheet containing all of the changes the agent made. So the MRP asked the Google Sheets Assistant to create a changelog spreadsheet.
The user also requested a spreadsheet containing all of the changes the agent made. So the MRP asked the Google Sheets Assistant to do that. It created a spreadsheet with three separate worksheets:
- A Change Log worksheet containing all of the changes made to the HubSpot CRM, including date, object, old value and new value.
- A stakeholders worksheet containing the contact details of all the stakeholders it found, and where the source was (already in HubSpot CRM, or through Apollo).
- An asset spreadsheet containing links to the slide deck it created, the Notion page, and the HubSpot portal with instruction for how to access the portal by providing your own portal ID. The best thing about this is that for Notion, it told the user that the link would be updated as soon as they provide the correct parent page ID so it could create the page. Totally removing any guesswork for the user on why the page wasn't created and how to resolve it.
Finally, the MRP agent delegated to Gmail assistant to send an email to the sales team with a summary of research and assets they need to help prep for the meeting.
Here is the email (the agent is using my account under the hood, ideally you'd give it its own account. We've done that in our prod version):
The last thing the MRP did, was summarise what it's original goals were, everything it did got done along with links to all of the assets, as well as direct actions points for the user to complete (provide Notion page ID), as well as suggestions for other things it could do if the user wants, like extra enrichment for each individual stakeholder.
You need to try this!
Thousands of businesses have already seen transformational results
Join the thousands of other teams empowering their workforce with AI.
Head of IT
Digital Ocean
AI Consultant and Youtuber
Founder and CEO
CRO
Founder and CEO
Connect your MRP Agents together into an MRP Workforce
