Dave Abalar outlines the five levels of AI agent complexity, emphasizing that most reliable production systems use simple, deterministic approaches like augmented LLMs, prompt chaining, and tool calling, while more advanced agentic systems (agent harnesses and multi-agent orchestration) offer flexibility but introduce greater risk and complexity. He advises engineers to start with the simplest solution and only increase complexity as needed, combining different approaches for optimal results.
In this video, Dave Abalar, an experienced AI engineer and founder of Data Luminina, discusses the five levels of AI agent complexity and what actually works in production environments. He emphasizes the importance of choosing the right level of complexity for your AI project, starting from the simplest solution and only increasing complexity as needed. Dave notes that while advanced agentic systems are exciting, most reliable production systems still rely heavily on simpler, more deterministic approaches.
The first three levels of complexity are covered quickly, as they are foundational and have been discussed in detail in previous videos. Level one is the augmented LLM, which involves making a single API call to a language model and using structured output. Level two introduces prompt chaining and routing using directed acyclic graphs (DAGs), which allow for deterministic decision-making based on input classification. Level three adds tool calling, where the LLM is given access to a set of tools to perform specific tasks, allowing for more flexibility and automation at the “edge nodes” of the system.
Dave explains that while DAGs remain the backbone of reliable B2B automation, they can become unwieldy and complex as projects scale. To address this, his team has started integrating tool-calling agents at the edges of these graphs, allowing the LLM to make decisions and call tools as needed. This hybrid approach combines the reliability and maintainability of DAGs with the flexibility of agentic systems, only resorting to tool calls when necessary to handle complex or ambiguous cases.
At level four, Dave introduces “agent harnesses,” such as OpenClaw and Cloud Code, which provide a runtime environment for agents to execute commands, access files, and interact with external APIs. These harnesses enable more powerful and autonomous agents but also introduce new risks and challenges, especially in production settings. Dave demonstrates how to set up such an agent using the Cloud Agent SDK, highlighting both the potential and the experimental nature of these advanced systems.
Finally, level five involves multi-agent orchestration, where multiple agents with separate context windows can be coordinated by an orchestrator. This allows for more scalable and modular solutions, particularly for complex, long-running tasks. Dave concludes by advising engineers to always start with the simplest solution and only add complexity as needed, combining different levels as appropriate. He also reflects on the current state of AI engineering, noting that while new tools like OpenClaw are exciting, they build on established principles, and understanding these fundamentals is key to navigating the rapidly evolving AI landscape.