An AI agent trace should show the run’s structure and causal sequence: which workflow ran, what the model did, which tools it called, and what retrieval or handoffs shaped the result. Capture timing, status, errors, and real correlation IDs by default; record prompts, tool arguments, and outputs only when there is a clear debugging or evaluation need and suitable privacy controls.
What an AI agent trace needs to explain
A useful trace lets you follow a run from orchestration through model inference, tool execution, retrieval, and any handoffs. Represent those operations as correlated spans or events, linked by parent-child relationships and ordered by time. This makes it possible to distinguish, for example, a slow model call from a slow tool or a repeated orchestration loop.
As an Amazon Associate I earn from qualifying purchases.
OpenTelemetry’s [GenAI agent span conventions](https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-agent-spans.md) describe agent and workflow spans for showing orchestration around individual operations. Google Cloud lists troubleshooting failed API requests, loops, and latency, validating communication flows, and evaluating output quality and cost among the uses of agent telemetry; these are documented product use cases, not independent study findings (Google Cloud documentation).
Recommended trace fields
| Trace component | Record by default | Capture only when justified |
|---|---|---|
| Run or workflow | Stable trace ID, workflow name, environment, start and end times, outcome, and parent/child links. | High-cardinality or implementation-specific detail that does not help identify the run. |
| Agent or workflow span | Agent/workflow identity, operation, timing, status, and parent-child links that reveal orchestration and handoffs. | Internal implementation-only invocations represented as separate user-facing workflows. |
| Model inference span | Provider or system and model identifier where available, operation, timing, status or error, and token/usage data where available. | Ordered system instructions, input messages, and model output. |
| Tool execution span | Tool name and type, call ID when available, timing, execution status, and errors. | Structured arguments and tool results, which may contain sensitive content. |
| Retrieval or context event | Relevant query or document identifiers/references and relevance scores where available. | Large retrieved documents or other sensitive context copied into telemetry. |
| Policy or evaluation event | Concise guardrail or evaluation outcome when needed to diagnose policy behavior. | Sensitive explanatory payloads behind the outcome. |
Use actual identifiers from the system. An existing conversation ID can help connect a run to its conversation, but do not invent a UUID, trace ID, or content hash and present it as a conversation identifier. Keep names low-cardinality where possible, and avoid duplicate spans for one tool call. OpenTelemetry’s conventions are actively maintained, and the GenAI conventions have moved to a separate repository; check the current attribute names and stability status before implementing them (OpenTelemetry GenAI spans).
#1 Best Overall
- EVOLUTION AMD RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Should you log prompts, tool calls, and context?
Prompts and model outputs
Usually, not by default. Prompts, user messages, and model outputs can be large or sensitive. OpenTelemetry’s semantic conventions say instrumentations “SHOULD NOT capture them by default, but SHOULD provide an option for users to opt in.” The conventions describe three approaches: omit content, optionally attach structured content to spans, or store content externally and put references in telemetry. External storage is described as a production option when payload volume or sensitive-data handling matters, since access controls can be managed separately (OpenTelemetry GenAI spans).
Enable content capture for a defined debugging, evaluation, or audit need—not simply because the instrumentation makes it easy. If you do capture content, decide who may view it, how long it remains available, and whether the same objective can be met with a reference or redacted excerpt.
Rank #2
- Built for Local AI Development: AMD Ryzen AI Halo is designed for local AI development and inference, featuring 128GB unified memory and support for up to 200B parameter models to build and run intensive AI workloads locally.
- 128GB Unified Memory: Features 128GB LPDDR5x unified memory at 8000 MT/s with 256 GB/s memory bandwidth, providing a shared memory pool across the CPU, GPU, and NPU to support larger AI models.
- AMD Ryzen AI Max+ 395 Processor: Features 16 cores, 32 threads, and Zen 5 architecture, paired with AMD Radeon 8060S integrated graphics featuring 40 RDNA 3.5 compute units and an AMD XDNA 2 NPU with up to 50 TOPS.
- Linux AI Developer Platform: Purpose-built for Linux-based AI development with full AMD ROCm software support and preloaded tools, models, and workflows optimized for local AI development.
- Compact, Connected Design: Includes a 2TB M.2 SSD, 10GbE LAN, Wi-Fi 7, Bluetooth 5.4, USB-C connectivity, and HDMI 2.1b.
Tool arguments and results
Tool-call metadata is useful even when the payload is not: record which tool ran, its call ID when available, when it started and ended, and whether it succeeded or failed. Treat arguments and results as opt-in content; they may expose personal data, confidential documents, credentials, or other sensitive information. Do not duplicate a single execution as multiple tool-call spans.
Free tools Windows power users keep installed
One-click scans. No signup required.
Retrieved context and handoffs
When context explains a model decision, represent it deliberately. Preserve message order when needed, and prefer retrieval identifiers or controlled references over copying whole documents into a trace. Record handoffs and their relationships so an operator can tell which agent or workflow passed work to the next one.
Rank #3
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Choose a capture design that fits the risk
There is no single right content-retention level for every environment. Compare designs across diagnostic value, privacy exposure, correlation quality, operating cost, and portability:
- Metadata-only: lower exposure and payload volume, but may not show why a model or tool produced an unexpected result.
- Selective content capture: opt in for specific environments, runs, or investigations; this can improve diagnosis while limiting routine collection.
- External content storage with trace references: keep telemetry lean and separate access controls for content, at the cost of managing another storage path.
OpenTelemetry provides a portable conventions starting point, while provider and framework fields may differ or evolve. Google Cloud and Microsoft Foundry document product-specific observability behavior; verify details against the service and deployment you use rather than assuming one provider’s defaults apply elsewhere.
Operational safeguards before enabling tracing
- Keep secrets out of payloads. Microsoft Foundry advises against storing secrets, credentials, or tokens in prompts or tool arguments. Prevent them at the source where possible, and apply redaction controls appropriate to your application.
- Restrict trace access. Give access only to people who need it, especially when traces can expose prompt or tool content. Microsoft documents access requirements for viewing traces; exact requirements depend on the deployment.
- Set sampling and retention intentionally. Use settings that match the environment and diagnostic need. Microsoft notes sampling and retention can be adjusted for cost management, and that retention is governed by workspace settings; pricing and retention vary by configuration (Microsoft Foundry tracing documentation).
- Check framework-specific defaults. The OpenAI Agents SDK documents tracing for generations, tool calls, handoffs, guardrails, and custom events. Its documentation says sensitive-data capture is enabled by default and can be disabled; that behavior is specific to the SDK, not a universal agent-framework default (OpenAI Agents SDK tracing).
A practical default
Start with correlated run, workflow, model, tool, and retrieval spans containing identity, sequence, timing, status, errors, and available usage metadata. Keep prompt and payload content off by default; add narrowly scoped capture or controlled external references when a defined debugging or evaluation task requires it. Revisit field names and stability as the OpenTelemetry GenAI conventions evolve.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
