Sign InOpen Brain
arXivPaperNeeds Review

TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development

TraceML exposes why coding agents stall on long ML work: they repeat narrow optimization loops instead of alternating work types, revisiting discarded ideas, and pivoting like experts.

arXiv · Aug 26, 2026
Open Source Open MarkdownOpen JSON
Source Summary

TraceML records **4,465 human Kaggle trajectories across 134 competitions**. Seven competitions also include **430 paired human and 207 agent trajectories**, with each code version labeled by action, intent, edit size, timing, score, and score effect.

Practical Implication

For long-running ML agents, plan explicit shifts among data, validation, modeling, and ensembling, and preserve discarded approaches for later review. Codex concentrated on ensemble weighting and submission tuning, while MLEvolve repeatedly mutated its current model.

Agent-Ready Context
TraceML records **4,465 human Kaggle trajectories across 134 competitions**. Seven competitions also include **430 paired human and 207 agent trajectories**, with each code version labeled by action, intent, edit size, timing, score, and score effect.

For long-running ML agents, plan explicit shifts among data, validation, modeling, and ensembling, and preserve discarded approaches for later review. Codex concentrated on ensemble weighting and submission tuning, while MLEvolve repeatedly mutated its current model.

A short planning prompt moved named behaviors toward the human profile and improved scores, but the agents' overall effort patterns remained distinct. The paired analysis covers only seven competitions and two scaffolds.
Connected Context · Feed7 Judgment

TraceML supplies a concrete human baseline for evaluating how ML agents allocate effort, revise plans, and preserve alternatives, rather than judging only final scores. The planning prompt shows that scaffolding can shift named behaviors and improve outcomes, but persistent differences from human effort patterns and coverage of only two scaffolds across seven paired competitions limit generalization.

From Agent Traces to Agent Simulations — Rustem Feyzkhanov, Snorkel AITraceML’s labeled trajectories provide the kind of replayable behavioral record needed to compare agent stacks under fixed conditions, extending simulation-based evaluation beyond final outcomes.Your Agent Evolved. Your Evals Didn't. — Ameya Bhatawdekar, BraintrustIts action, intent, timing, and score-effect labels operationalize trajectory-level coverage for planning agents, supporting the claim that final-answer evals no longer cover richer architectures.Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayerThe observed gap between agent and human planning reinforces the need for human ownership and upfront alignment; higher task scores alone do not establish that the development process remains reviewable or maintainable.
Context Map
benchmarkcodingdata#agent-evals#agent-reliability#harness-engineering
Uncertainty
A short planning prompt moved named behaviors toward the human profile and improved scores, but the agents' overall effort patterns remained distinct. The paired analysis covers only seven competitions and two scaffolds.