Tracing
Tracing records a program's execution to find errors or performance issues.
Tracing is like watching a detective follow clues to solve a mystery. In AI, it means recording every step a program takes as it runs. This helps developers see exactly what the AI is doing, find mistakes, or figure out why it's slow.
For example, if a voice agent misunderstands a command, tracing can show the exact audio processing steps and decision points that led to the error. This detailed log allows developers to pinpoint the problem and fix it.
Tracing is a debugging and performance analysis technique that captures the execution flow of a program. It involves instrumenting code to log events, function calls, and variable states as they occur. This detailed record, often called a trace, allows developers to reconstruct the program's behavior, identify bugs, and diagnose performance bottlenecks.
Tradeoffs include the overhead introduced by tracing, which can alter program behavior and performance, and the potentially large volume of data generated, requiring efficient storage and analysis tools. Different tracing frameworks offer varying levels of detail and performance impact.
Tracing is becoming more important as AI systems grow larger and more complex. The Anthropic and Microsoft examples show how companies are using tracing to understand and improve the performance of their AI applications, from API calls to voice agents.
- Anthropic Cookbook: feat(observability): add OpenTelemetry tracing cookbook (#922)Anthropic · Anthropic Cookbook ·
- Build, test, monitor and optimize a voice agent in Microsoft FoundryMicrosoft · Microsoft Azure on YouTube ·
- Anthropic Claude
- Anthropic is adding OpenTelemetry tracing tools to help users track their Claude API calls.
- Microsoft Copilot
- Microsoft is using tracing as part of its tools for building and monitoring voice agents.
- Profiling
- Profiling focuses on measuring the performance of code, like how much
- Logging
- Logging typically records discrete events or messages, whereas tracing
- Debugging
- Debugging is the general process of finding and fixing errors; tracin
Tracing as a debugging technique has roots in early computer science, evolving from simple print statements to sophisticated instrumentation tools. Modern tracing systems are often built upon established principles of observability.
You can see a form of tracing by using the developer console in your web browser. When visiting a website, open the console and look at the 'Network' tab. This shows you the sequence of requests the browser makes to load the page, similar to how tracing tracks program steps.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- WorkflowBuild, test, and monitor a voice agent in Microsoft FoundryEnd-to-end process for creating, customizing, testing, and optimizing voice agents using Azure Speech. · 120 min
- PromptRole-play a student struggling with math for AI tutor simulationSimulate a student's math struggles and uncertainty for AI tutor evaluation. · 5 min
- TemplateReusable workflow templates for healthcare teamsTeams can create and share templates for common workflows like documentation or prior authorizations. · 15 min
- TemplateTemplate for building an AI tool tailored to your workA structured approach to defining and building a custom AI tool for specific business needs. · 15 min
- TemplateEvals CSV prompt list for agent skillsA CSV file template for listing prompts used in evaluating agent skills. · 20 min
Questions people ask
- What is the difference between tracing and logging?
- Logging records specific events or messages, like "user logged in." Tracing records the entire flow of execution, showing how the program got from one state to another, including function calls and variable changes.
- How does tracing help with AI?
- Tracing helps AI developers understand the complex decision-making process of AI models and agents. It allows them to identify where errors occur or why an AI might be behaving unexpectedly.
First seen on Subvolts Sep 28, 2026. Updated Oct 14, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.