After the release of ChatGPT in late November 2022, the race toward general artificial intelligence (AGI) and the development of foundational LLMs dominated media headlines for a good 18 months. In 2024, that narrative has subsided. People have started to grow wary of generative AI’s ability to add business value. Foundational models felt like they could do everything and nothing at the same time. They had their wow moment but then what? How can you apply such broad technology to businesses that have very specific problems?

At Squad, we believe the answer to that question is agentic AI. To put it simply, an AI agent is a program designed to perform (or have “agency” over) a specific task or set of tasks. An agentic system, like Squad, is where multiple AI agents work together to achieve a more complicated task or set of tasks.

The idea of having agents work for you isn’t new. It’s a commonly used paradigm when working through bigger, multi-faceted problems. Let’s look at an example. When someone wants to buy a house there is a problem—the buyer wants to buy the house for as little as possible and the seller wants to sell the house for as much as possible. So how do you solve this? You break the problem down and use real estate agents. These agents are given “agency” to act on behalf of the buyer and seller and reach the best possible outcome for their clients.

What does that have to do with product management?

People building products also face big, multi-faceted problems. Rising customer expectations & growing customer acquisition costs (CAC) have steered companies towards enhancing product experience to deliver on business outcomes. But developing a product to both deliver on business outcomes while meeting the expectations of its users is no simple feat. Product managers must learn what their customers want, what their business needs, how their product is performing, identify what they can build to improve it, create a clear product vision, and then work with multi-disciplinary teams to actually implement it.

Product managers have tried to break this problem down and have developed endless methodologies and processes for tackling it. Unfortunately, these methods are often very rigorous, need lots of interaction with users, and require large teams that are aligned on what they’re building and why. When you consider that the software development industry is becoming more and more accessible, these methods are no way achievable for the majority of people actually building products today.

Yet, this is exactly the type of problem that agentic AI can solve. A big, multi-faceted problem that has been broken down into smaller processes in which agents can help automate. This presents an opportunity that not only provides a more accessible solution to the original problem, but still allows for control and user involvement at specific points in the process should a user want it.

Looking at how we apply agents to these smaller processes, we can break down the first part of the product development process (product strategy) into 4 phases:

Each one of these phases can have its own specialised agent, focused on helping understand and complete tasks such as:

Discovery agent: What does the business need? What do our customers want?
Ideation agent: What could we build? What can we build?
Definition agent: How should we build it? What is the effort to build it?
Planning agent: When should we build it? How should we prioritise it?

Each of these agents can then work together to form an agentic system for product strategy, following the same tried and true processes of a mature and effective product team.

How do agents act autonomously and work together?

One of the early challenges we faced at Squad is how to solve for continuity of knowledge between agents, i.e. how they work together or form a system. We solved this in two ways:

1. We allow the agents to be controlled by a master orchestration agent. The orchestration agent hosts foundational knowledge (a knowledge base) about the product and process as a whole to make sure each agent is given context and completes its tasks more accurately. This process is done through a knowledge base and retrieval augmented generation (RAG) by each agent.

2. We centre the entire product development process around an outcomes-based roadmap or Opportunity Solutions Tree (OST). This is beneficial for the users of Squad as it keeps their product strategy aligned to business and customer needs, but also—very conveniently—it follows a similar data structure to that of a graph database. The graph data base allows for a multi-relational data structure and therefore acts as the network of thought contributed to by each of the agents as they work through their individual tasks.

Its the combination of these architectural decisions that result in what is known as “AI native” software and, more specifically for Squad, an operational model for agentic product development:

Breaking this down further, let’s look at the Discovery agent which is tasked with answering:

What does the business need?
What do our customers want?

In the exact same way as a product manager, this agent needs to take in strategic decision making data such as: customer feedback, product analytics, market research, and existing knowledge of what is being built; to understand a product’s current state of play and provide useful information to the Ideation agent. Squad’s discovery agent does this—backed by the orchestration agent and existing data in the database—by ingesting this data, modelling it into categories and topics, and then deriving any relevant insights from it. This is packaged up in the form of “Opportunities” (an opportunity for a product team to improve or action against) which is contributed to the underlying graph data base and passed onto the ideation agent to start creating ideas for what a product team could build.

Admittedly, this is a simplified description for the sake of the article—one that will likely get me in trouble with our engineering team. There will always be nuances with each agent, data standardisation can pose a challenge and the model of the underlying graph database can quickly grow in complexity. The process and underlying data structure also needs to consider human augmentation and provide a UX that allows for a user to insert, delete, realign, stop, or regenerate information at relevant points in the process.

So, where are we heading?

We’ve provided an example of taking a problem, breaking it down into smaller steps or processes, and creating a system of agents to work together to solve it. Where this gets really exciting is when you zoom out further. Another huge benefit and something we believe will become prevalent as more of these systems are built, is that agentic systems like Squad are also very suited to working with other agentic systems. In the case of product development, AI is already very apparent in the other disciplines involved in building software. It’s not difficult to imagine a world where agentic systems from each function work together to automate the entire product development process—from product strategy right down to quality assurance engineers.

This is why the use case for AI goes far beyond chat bots, it creates an entirely new operating system in which software can be built. For product management, that starts with AI-native tools like Squad.