Multi-Step Behavior, Instrumentation, and Streaming

Lecture 6.4

• Demo: multi-step tasks the agent chains autonomously (read → edit → verify)
• Emergent behavior: the agent plans steps it was never explicitly told
• When the agent gets stuck or loops — recognizing and debugging failure modes
• Instrumenting context growth: counting tokens per API call, tracking the messages array size over a session
• Visualizing the cost curve: where tokens accumulate and why
• Streaming responses: implementing streaming for better UX during long tool-calling sequences
• Connection to Module 4: context growth is the problem, context management is the solution

Additional Resources

• Lecture slides