TraceDance: Automating the Construction of Agent Behavior Benchmarks from Real-World Deployment Traces

Meta Description: TraceDance builds 107 agent benchmarks from 252,557 deployment sessions and finds nine frontier LLMs pass only 26.7% of targeted behavior tests.

Tags: TraceDance, agent benchmarks, LLM evaluation, deployment traces, tool use

Featured Image: Abstract technology style: a flow of session logs transforming into a benchmark grid, no real people, no company logos, no text.

TraceDance is an agent system that converts deployment traces into targeted benchmarks for undesirable behaviors observed during execution. It matters because fixed benchmark suites often miss the specific ways agents fail in coding and general tool use, while developers need tests for the behaviors they encounter in deployment.

From fixed tests to deployment traces

An agent can finish a task and still exhibit an undesirable behavior while doing it. Traditional benchmark suites can measure final outcomes, but they do not always capture the narrow, operational failures that appear after an agent leaves the lab. TraceDance starts from that gap. Its premise is simple: if a behavior is worth avoiding in deployment, it should become a test case.

That framing shifts the evaluation problem from asking only whether an agent reached the right result to asking whether it behaved in a specific, unwanted way at a particular moment. For developers building agents around coding and general tool use, that distinction can matter more than a single aggregate score.

Building benchmarks with TraceDance

The system turns deployment traces into benchmarks for user-specified undesirable behaviors. To make that process efficient, TraceDance uses Anchor-and-Confirm, which combines programmable retrieval with candidate-level confirmation by a Flash large language model (LLM). That means the system first finds likely trace segments, then asks the Flash LLM to confirm whether a candidate actually shows the behavior the user wants to test.

For custom behaviors, the Anchor Synthesis Loop generates and revises specifications. The loop is needed because deployment problems are not always described with ready-made test labels. A developer may know an agent should not take a certain action, but the exact test criteria may need to be formed from real examples.

The resulting benchmarks use decision-point continuation. At a recorded decision point in a trace, the system evaluates an LLM's next turn against a behavior-specific rubric. This avoids requiring a reference answer or environment replay, which can be costly or impossible for complex agent sessions.

Results across 252,557 sessions

Experiments in coding and general tool use draw on 252,557 sessions and produce 107 benchmarks with 4,125 instances. The process fulfilled 95.3% of build-target requests, showing that most requested behavior tests could be constructed from the available traces.

Quality checks support the method. Both human annotators confirmed the requested behavior in 84% of sampled instances. The automated grader's agreement with human pass/fail judgments was comparable to the agreement between the annotators. That matters because the value of a deployment-derived benchmark depends on whether its labels reflect real behavior, not just a convenient extraction rule.

The model results are the sharper signal. Nine frontier LLMs achieved a mean pass rate of only 26.7%. The number does not prove that the models cannot complete tasks; it shows they often fail the next-turn behavior test at the evaluated decision point. Analysis across behavior-specific benchmarks further reveals weaknesses in how current LLMs behave as agents.

What this stage means

TraceDance is best read as a method for making agent evaluation more operational. It does not replace broad capability benchmarks; it adds a way to convert observed deployment problems into repeatable tests. That can give teams a clearer record of which agent behaviors are still fragile, especially in coding and tool-use settings.

The recursive self-improvement (RSI) angle follows from that role: if AI systems can help build better AI systems, targeted benchmarks derived from real traces can become part of the feedback loop. The practical question now is whether trace-based benchmarks are adopted widely enough to change how agent failures are tracked, and whether frontier LLM pass rates on behavior-specific tests improve as deployment evidence becomes part of evaluation.