There are many posts providing insights about Agentic AI, the protocols (MCP, A2A, etc.), the frameworks (Langchain, Google ADK, etc.), and not to mention how amazing AI Agents are because they have ‘memory’ and ‘actions’.
What no one talks about is the ‘how’. How does one build and operate agents and multi-agent systems?
This lack of the ‘how’ is what leads to expectation mismatch and the selection of agentic solutions where a simpler solution would have delivered value quicker, with less effort and cost.
In this post I want to talk about the ‘how’.
The First Step
The first step is to get your approach correct. Multi-agent systems are less like building IT Apps or Gen AI Apps and more like building a team of people. You have to continuously iterate between the individual and the team because a change in one will impact the other.
Start defining the agent-enabled journey in terms of:
- expected outcomes of the journey
- tasks associated with different stages of the journey
- constraints (hard and soft) associated with the tasks
- tools required for the tasks
- communication pathways between tasks
- ontology/knowledge/data to carry out the task
- handoffs and human-in-the-loop mechanisms for each task
- information exchange mechanisms with external entities (e.g., other systems and humans)
Don’t start by worrying about Agentic frameworks, MCP, A2A, etc. These will help you build correctly, not build the correct thing.
The Next Step
Go to the next level of detail:
- How do the tasks, constraints, tools and knowledge group into agents? This is not about writing code. Coding complexity of agents is low. For AI agents complexity is in writing the prompts and testing.
- Describe how will we test individual agents in isolation then how do we start bringing them together. Can agents deal with failures around them (in other agents)? Can they deal with internal failures and degrade gracefully?
- How will we monitor the agents? What patterns are we going to use (e.g., Watchdog pattern) to enable monitoring without degrading agent performance? What actions can be taken to deal with issues identified. Can the agent be isolated rapid to prevent scaling up of the issue? Think of this like writing a diary at the end of the day where you describe and rationalise what you did. Relevant to the agent, interesting perhaps for other agents in the group.
- How will we test Agent ensembles?Validate parts of the multi-agent system by describing inputs to each agent, outputs provided, failure scenarios, and upstream/downstream agents. How do groups of agents deal with issues? Can they recover or at least prevent failure from spreading beyond their boundaries? Can they prevent external events (to the ensemble) from disrupting the ensemble?
- How do we monitor agent ensembles? How do we combine streams from related agents to give a view of what these agents are up to. Remember with agents grouped together we will need to stitch a narrative from the monitoring feeds. Think of it like folktales relevant to few agents within a group but interesting for other related groups to know.
- Bring Agent ensembles together and start to test the whole system. Validate inputs to the system and expected outputs, failure scenarios, and external entities the system interacts with. The exact same layering as with the ensemble with the same set of questions but answered at a higher level.
- How do we monitor the whole system. Remember the whole system includes the operators, users, agents, IT systems, and the knowledge required and generated. So monitoring needs to feed the system-wide narrative of what is going on. Think of this like a history of a civilisation. All about what agents/users did to get us to this point. Relevant for everyone.
Hopefully by this time you are convinced of this layered step-by-step approach. How individual interactions give rise to interactions between groups and so on. The same scaling works for other aspects like testing, monitoring, recovery, and operations.
Finally, hope you are excited about the journey that awaits you as you enter the world of Agents!
Fish Eye View 