I can’t wait until I start packaging this Multi-agent Communications Framework, or my “AI employees” as you could call it, that I’ve built (it uses Cursor IDE, at least for now). As a prompt engineer, one of the guiding principles is to ask an AI LLM to “act” as a specific role, and the more it can act as and the more knowledge it has as that role, usually the better results you can get. The idea behind this Multi-agent Communication Framework (as I’m calling it now at least) is to have one AI automatically break up tasks by role, and act as multiple roles, each talking back and forth with each other so I don’t have to do this over and over manually between agentic roles.
It has been a bit of a process to figure out the quirks and also train the individual AI “employees” to be good in their individual roles, but as I add tools and knowledge at their disposal, with the power of AI and software automation I almost have a machine of AI employees that coordinate with each other, plan with each other, delegate to each other all with the power and ability to write software on their own and manipulate file systems and even work remotely if needed.
The trick has been using many of the Agile software development and product management techniques taught to me by Alistair Cockburn (along with a few others throughout my career), who was also one of the creators of the Agile Manifesto which kicked off Agile Project Management as a methodology. The idea behind Agile is to reduce overhead and process in project management and break it up into smaller chunks that can actually be measured and the results “seen”. This reduces “scope creep” and ensures that every step of the way you’re only promising what can be actually measured and seen by the project team.
I start each project with a Release Planning “meeting” with my AI agent team of all relevant roles to that project, floating ideas off of the team, assigning a Scrum Master (which may become its own specific role in the future). We create a backlog of user stories, all intended to meet the requirements and goals of the project at hand and the desires of all stakeholders (sometimes just me, sometimes it could be a client).
Then we start the first sprint planning meeting where I specify some basic goals I want to achieve first within the release plan and I have “the team” pick a list of user stories from the backlog to complete during the sprint. They even assign weights to each user story to be sure we’re not taking on too much at a time.
Then comes the powerful part, and whether we do it in sprints or it’s just me asking my Executive Secretary/Personal Assistant to do something, this HAS to happen so there are measurable results for everything the AI team does: we establish a “definition of done” and then the team creates measurable and testable acceptance criteria to define what gets us to that definition of done and that gets added to each user story for the sprint.
Here’s the cool part: once those are in place we can apply actual software testing, with AI, to measure if each acceptance criteria is complete and when the user story is satisfied. We then build scripts through Behavior and Test-driven software development principles (using libraries like behat and PHPUnit) that get added to a regression test and sometimes GitHub workflows to AUTOMATICALLY define when a project is complete.
A lot of agentic AI developers talk about what they call, “vibing”. This is when, for a larger software development project, the AI kind of goes in circles with endless frustration from the developer just trying to make AI create the end project.
My approach takes a measured agile approach to force AI, working as a team of experts each with roles to fulfill, that can know exactly where they stand every step of the way and accomplish not just software projects, but managing just about anything you want to organize in life!
It still veers off here and there but the more in work on it and as soon as I get it in a more controlled environment like an extension, this will be a well-oiled machine with perhaps even a database of trained roles you can purchase. For instance, I’m currently trying to train my Graphic Designer role to use various image manipulation APIs and image creation AIs as well to do things like create professional logos and entire brand design documents, automatically. And I’m almost there!
If this piques your interest, I’m open to sharing it with others willing to be patient with it with my help as I make this a more solid product. It will be around $5k with a few months of support due to my need to hold your hand as you get started but we can also work something out where you get the packaged product for a given time once I have it ready. Reach out! I’m serious at how valuable this truly is!
In the meantime, stay tuned to hear about the first product I built with this that I’m launching soon that some of you in Austin, Texas might like! That’s what I’m training the designer to brand for right now. 😉
