In the evolving world of artificial intelligence, the development of AI agents often faces inefficiencies and mismanagement. Ara Khan, in a recent presentation titled "Don't Build Slop (4 Levels of AI Agent Maturity)" at AI Engineer Europe, emphasized the need for a structured framework to help developers avoid common pitfalls and improve agent design.
Khan's talk, sponsored by prominent companies including Google DeepMind, Braintrust, and WorkOS, highlighted the significant issue of "slop" in AI agent development. He noted that the complexity of coordinating multiple agents can overwhelm even experienced engineers, resulting in substantial productivity losses. Two main challenges identified were inference bounds and take isolations. Inference bounds occur when developers must wait for inference processes to finish, while take isolations arise when multiple agents trained on the same data create conflicts during integration.
The Four Levels of AI Agent Maturity
To tackle these challenges, Khan introduced a four-level model aimed at enhancing AI agent maturity:
Level 1: Use a Framework
In this foundational stage, developers are encouraged to use existing frameworks such as LangChain, LangGraph, CrewAI, AutoGen, or LlamaIndex. This level focuses on grasping the basic architecture and functionality of AI agents, setting the stage for further development.
Level 2: Build It Yourself Agents
The second level urges developers to create agents from scratch. This stage highlights the importance of architecture, modularity, and model independence. By implementing state machines—recursive loops with defined conditions and end states—developers can better understand agent behavior.
Level 3: The Kanban: Visualize the Agent's Flow
At this stage, the emphasis shifts to visual management of agent activity. Developers are encouraged to create a clear overview of each agent's progress, reducing context switching. This level introduces dependency chains that enable tasks to be linked autonomously, boosting operational efficiency. Features like "diff review on click" allow developers to track changes as agents complete their tasks.
https://www.youtube.com/watch?v=yUmS-F9IX90
Level 4: Ship It to the Cloud: Parallelized, Production-Grade
The final level focuses on deploying agents as production-ready, cloud-native systems. Key features at this stage include the ability to run multiple instances in parallel, removing local dependencies, and using APIs for programmatic orchestration. This level also supports horizontal scaling, enabling the deployment of more agents rather than just larger ones.
As the AI field continues to expand, adopting this maturity framework could greatly enhance how developers approach AI agent creation. By concentrating on these four levels, teams can manage the complexities of multi-agent orchestration, leading to more reliable and efficient AI systems. Khan's insights serve as a crucial reminder that each addition to an agent risks complicating its functionality. He aptly stated, "Every single thing you add to an agent risks making it worse." With this framework, the path to more mature AI agents can become more structured and beneficial for the industry.
Implications for the Future
Embracing Khan's proposed framework may not only streamline AI agent development but could also drive breakthroughs in various applications across industries. As more developers adopt these structured methodologies, the potential for creating sophisticated, scalable, and efficient AI solutions is likely to grow, pushing the market forward into 2026 and beyond.
The stories that move AI & crypto markets — before the market reacts.
Free. 7am ET. Five stories. 62,400 readers.
