This chapter is from the book
This chapter is from the book
This chapter is from the book
Conclusion
The leap from workflows to agents to multi-agent systems can unlock a whole new set of capabilities for AI applications. Workflows are great choices when you want efficiency, predictability, and full control. Agents bring adaptability, flexibility, and a little bit of creative chaos, which can be a wonderful thing in some cases.
In this chapter, we broke down not just how to build these systems, but also how to evaluate them using both human-driven rubrics and automated tools like LangSmith for traceability and auditability. Although agents can sometimes cost more and take longer, they also offer room for learning, collaboration, and improvement over time. That is especially the case when the agent is given the ability to write, recall, and share evidence with itself (recall our “Otto” example) or with other agents (as in the SDR framework).
Bottom line: There’s no single “best” approach to developing systems with AI agents. The right choice depends on your use-case, your tolerance for ambiguity, and how much control you want versus how much flexibility you need. Try both approaches. Experiment, evaluate, and build a system that works for you, and don’t be afraid to combine the best parts of each.
In Chapter 5, we’ll do just that. We’ll create agentic/workflow hybrids and then take these ideas even further, exploring long-running, multi-agent workflows that can handle complex, evolving tasks over time. See you there!
