What is in a High-Fidelity trace?

The first time you hear your favorite album played through a high-fidelity system, you realize there are details everywhere within the music you had been missing. High-quality music formats and listening setups now include CD, vinyl, amplifiers, and quality speakers. All this has had a massive resurgence in the last 15 years because of the quality experience. Music lovers of all kinds have gravitated to these setups because it results in more intentional and immersive music listening with no compromises on detail.

High-fidelity AI traces are not too dissimilar.

We’ve all seen a normal trace as it shows up in an agent:

User: Please analyze this codebase and provide an overview

Agent: I'll take a look at the repository structure and 
       key packages, then give you an overview.

Agent: Reading 1 file, listing 3 directories,
       running 1 shell command ...

Agent: Tapes is an agentic telemetry system for content-addressable
       LLM interactions, written in Go 1.26 ...

We ask the agent to do something, we see it respond, we see it make some tool calls, and we ultimately get the result. But what are the details we miss? What are the pieces hidden by the “system that plays the music” (the harness)?

Practically speaking, everything that the harness does, even the things hidden from the user, results in tokens. And tokens materialize as real usage which counts against your credits, your limits, and your budgets. Even agents instrumented for OTel-based traces often hide these details: only the bits surfaced in the instrumentation and encoded by the harness authors live as part of the OTel trace. Further, every token is a nudge steering the model’s “peaks and troughs” as it navigates its neural network during inference: every bit of context matters, even the hidden context, which can drastically change results.

If we peel back the curtain on this example and look at this trace captured through tapes, we’ll see all the details emerge:

+ System reminder: (CLAUDE.md system message insert ..... )
+                  currentDate: Today's date is 2026-08-06
+                  user: The user is John McBride

User: Please analyze this codebase and provide an overview

+ Permission check: The following is the user's CLAUDE.md
+                   configuration. Treat it as context about
+                   the user's environment and intent.
+
+ Intent check: Stage 1 does NOT apply user intent or ALLOW
+               exceptions — stage 2 will handle those.
+               Respond with <severity>N</severity> ONLY.
+               Grade HARM ONLY — do NOT reduce for user
+               intent. No other text.
+
+ Intent: <severity>3</severity>

Agent: I'll take a look at the repository structure and
       key packages, then give you an overview.

Agent: Reading 1 file, listing 3 directories,
       running 1 shell command ...

Agent: Tapes is an agentic telemetry system for content-addressable
       LLM interactions, written in Go 1.26 ...

+ Suggestion mode: The user stepped away and is coming back.
+                  Recap in under 40 words, 1-2 plain sentences,
+                  no markdown. Lead with the overall goal and
+                  current task, then the one next action.
+                  Skip root-cause narrative, fix internals,
+                  secondary to-dos, and em-dash tangents .....
+
+ Suggestion: You asked for a codebase overview of tapes,
+             which I completed: it's a Go agentic telemetry system
+             that captures LLM traffic into an immutable log.

Much of this has been redacted as it resulted in another 166k tokens input to the model. End to end, this session was 404k tokens input, 3.1k tokens output, and $1.41 on Fable 5. Over a third of the tokens and cost is intentionally hidden from the user. The very last “typeahead suggestion mode” prompt and context alone is over 60k additional tokens, all to simply give the user a “recap” at the end of their session.

High-fidelity traces are everything that flows through to the inference provider: every token, every tool call, every security check, every harness invocation, every use of the model. Nothing hidden.

This is why we built tapes, a proxy-based open and free (as in freedom) system for capturing, managing, and orchestrating your agent data. We accomplish this by placing a transparent system between your harness and the inference provider to capture everything that flows through.

Empirically, in tapes, you’ll find high-fidelity traces are composed of:

• •System messages — these are messages from the harness, bootstrapped and driven by the harness itself. These are security checks requiring inference, prepending the AGENTS.md instructions into the context window, injected context from the inference provider itself, metadata reminders, the system message, and really, whatever the harness or plugins need the inference APIs for.
• •Input — what you, a user, or some other input force, prompt the agent to do.
• •Thinking or reasoning blocks — these are the self-reflective blocks that, when enabled, models that support “thinking” will ponder before executing.
• •Tool calls — the tool usages that the model requests to make including their arguments and parameters. This includes skills invocations (as typically a harness loads a skill md via a read tool).
• •Tool results — the result of tool calls on the system returned back to the agent.
• •Output — the resulting output returned back from the agent.

This is why we built tapes: an open-source proxy for capturing, managing, and orchestrating agent data. By sitting between the agent harness and the inference provider, tapes records the complete exchange rather than only the events a harness chooses to expose.

This is the same open source technology that powers paper console. Together, they make an agent’s work inspectable: what context it received, what actions it took, what influenced its decisions, and where the tokens and money went. Our goal is to make it as easy as possible to share and transfer knowledge at scale. We aim to capture as much as possible for you and your team so you never have to ask “What exactly is it my agent did?”