Overview of Taking a look under your agent’s hood
In this episode of the Stack Overflow Podcast, Ryan Donovan speaks with Yanbing Li, Chief Product Officer at Datadog, about how observability is evolving for AI agents. The conversation focuses on what it means to “look under the hood” of an agent: tracing prompts, tool calls, reasoning steps, outputs, failures, and downstream effects across software systems. Yanbing explains that while the core observability concepts remain familiar, AI introduces new challenges like non-determinism, rapid iteration, security risks, and the need to connect production telemetry directly back into development and evaluation workflows.
Key Topics Discussed
From traditional observability to agent observability
- Traditional observability pillars like logs, metrics, traces, and events still matter.
- AI agents add new layers:
- prompts and responses
- reasoning paths
- tool usage
- decision-making behavior
- application-level outcomes
- The shift is from monitoring software performance to understanding software behavior.
Why AI agents are different
- AI systems are non-deterministic, so they can’t always be validated like traditional software.
- Agents operate as distributed systems, but with an added “intelligence” layer from models.
- Common issues include:
- API rate limits
- tool-call errors
- data access problems
- PII exposure
- incorrect or unsafe decisions
Observability as a continuous feedback loop
- For AI, observability is not just for production monitoring.
- It should feed back into:
- evaluation
- experimentation
- test generation
- model selection
- prompt/tool refinement
- Yanbing emphasizes that continuous monitoring and rapid iteration are more important in AI than in traditional software.
AI in the SDLC
- Datadog is thinking about how AI can automate more of the software development lifecycle:
- testing
- code review
- merging
- deployment
- rollout
- The goal is to make the loop from build → observe → improve more autonomous.
- Observability data can help generate:
- synthetic tests
- new test cases for emerging behavior
- production-informed validation earlier in development
Local development, MCP, and access to production context
- There’s growing interest in bringing observability into the developer workflow:
- IDEs
- CLI tools
- Slack
- local coding agents
- Datadog is experimenting with giving developers production context through MCP so agents can query observability data more intelligently.
- Yanbing notes that MCP should be more than a simple API wrapper; formatting, error handling, and smart querying matter.
Dashboards are not going away
- AI may change how people consume observability data, but dashboards still have value.
- Humans often want:
- trusted visual summaries
- fast, familiar access
- a way to validate what AI found
- At the same time, AI agents can:
- investigate incidents faster
- synthesize context across systems
- generate dashboards on demand for specific investigations
Observability, security, and cost are converging
- Yanbing sees strong overlap between observability and security:
- understanding every agent action helps with both monitoring and defense
- He also highlights the emerging need to track:
- token usage
- model efficiency
- ROI of AI systems
- These areas are not fully solved yet, but they all depend on understanding agent behavior deeply.
Main Takeaways
- AI observability is broader than model tracing — it includes full agent trajectories and system-wide effects.
- Feedback loops matter more in AI because testing can’t fully cover real-world behavior.
- Observability and development are converging as production data increasingly informs build-time decisions.
- Dashboards remain useful, but AI is changing where and how observability data is consumed.
- Security and cost optimization are becoming part of the observability conversation, not separate concerns.
Notable Insights
- “We started with observing the LLM, but now we need to understand the AI agent as a whole.”
- “The shift is from observing software state to observing software behavior.”
- “Agent observability is no longer just monitoring production; it’s part of the flywheel that improves upstream development.”
- “The boundaries are blurring between developers and SREs in AI-driven systems.”
Conclusion
This episode frames AI agents as the next major observability challenge: not just systems to monitor, but systems whose behavior must be continuously evaluated, explained, and improved. Yanbing Li’s perspective from Datadog suggests that the future of observability will be tightly linked to agent development, autonomous testing, security, and operational decision-making—with humans still in the loop, but increasingly assisted by AI.