Agent Evaluation · Benchmark Methodology

The benchmark score is also
a harness score.

First surfaced in Tandemly Briefing — 2026-06-13.

Researchers at Beijing University of Posts and Telecommunications, Shanghai AI Lab, and collaborating institutions ran 15 LLMs across 7 widely-used agent benchmarks. They used a single standardized harness for every model and every benchmark. What they found: published scores contain two variables mixed together. The model is one of them. The scaffold is the other.

Core concept
When every benchmark ships its own harness, a score reflects both the model and the harness it was paired with. Separating those two variables requires running every model through one harness. The paper builds that harness and shows how much it matters.
scroll to explore

Every benchmark ships
its own scaffold.

That means every published score is actually a score of model plus scaffold. Researchers and teams making decisions from those scores are reading a compound number as if it were a simple one.

AI agent benchmarks work like this: a team publishes a dataset of tasks, writes the infrastructure to run models against those tasks, and reports how well various models perform. The benchmark covers things like web navigation, code execution, multi-step reasoning, or tool use. It becomes a standard comparison point.

The problem is that the infrastructure, the harness, is not neutral. Each benchmark team makes choices about how the agent is prompted, how tool calls are formatted, whether the environment is reset between attempts, how failures are detected, and how partial completions are scored. These choices are engineering decisions, not science ones. And they interact with model behavior in ways the benchmark team did not necessarily intend or test.

The result: when model A outperforms model B on benchmark X, that score could mean model A is genuinely better at the task benchmark X was designed to measure. Or it could mean model A is better-matched to the particular scaffold benchmark X uses. Without a controlled comparison, you cannot tell which.

For production teams choosing between models or evaluating their own agent stack, this is more than a methodological inconvenience. It means that two benchmarks showing different rankings for the same models could both be correct, and still tell you very little about which model performs better in your specific setup.

The question this paper asks

If you run many models through many benchmarks but control the harness across all of them, how much do the rankings change? And can you separate the model's contribution from the harness's contribution cleanly enough to be useful?

One harness.
Seven benchmarks.
Fifteen models.

The team built a unified evaluation framework that converts seven different agent benchmarks into a single instruction-tool-environment format, then runs every model through a fixed ReAct loop. The benchmark's original scoring criteria are preserved; the scaffold is not.

ReAct (Reasoning plus Acting) is a standard loop structure for agent evaluation: the model receives a task, reasons about what to do, takes an action via a tool, observes the result, and reasons again. Different benchmarks implement this loop differently. The unified framework standardizes the loop while leaving each benchmark's underlying task set and success criteria intact.

Seven benchmarks were adapted to this format, covering 24 domains. The selection spans single-agent tasks, multi-agent coordination, and safety-critical scenarios, so the framework is not optimized for any one class of problem. Fifteen models were evaluated across these benchmarks using the unified harness.

The framework also includes an offline snapshot mode. In ordinary evaluation, some benchmarks involve live environments that change between runs: web pages that update, APIs that go down, services that return different results. This volatility contaminates comparisons because it introduces a source of variance that has nothing to do with the model. The offline mode replaces live environments with curated, stable snapshots, letting the researchers isolate environment effects from model effects.

The key design choices

The framework separates two things prior benchmarks merged: the task specification (what success looks like) from the infrastructure (how the model interacts with the environment). It keeps the former intact and replaces the latter. This lets the researchers attribute performance differences to a single variable at a time.

Unified resource metrics (token usage, latency, cost) and failure attribution were added across all benchmarks. Previously, comparing resource consumption across benchmarks required reconciling different reporting conventions. The framework makes these comparisons direct.

400K
Total rollouts
5B
Tokens consumed
15
Models evaluated

Scaffold choice moves
outcomes both ways.

The headline finding is not that harnesses matter more than models. It is that they are not separable in published evaluations, and the gap is wide enough to reorder leaderboards when you control for one and vary the other.

What published benchmarks show
Model scores vary across benchmarks. Different benchmarks give different rankings for the same models. The conventional explanation: benchmarks measure different capabilities. So the variance is informative.
What this paper shows
Scaffold choice and environment volatility explain a significant portion of that variance. Under a fixed harness and stable environments, the model ordering changes. The leaderboard is not measuring what it appeared to measure.
Finding 1: The harness affects outcomes in both directions

When the same model is evaluated through different harnesses on the same task set, performance can shift up or down. This is not a small correction. It is large enough to change which model ranks first. A model that appears dominant on a benchmark using that benchmark's native scaffold may rank lower under a different harness, and vice versa.

This is not a claim that one harness is better than another. It is a claim that scaffold choice is a variable, and published evaluations do not treat it as one.

Finding 2: Environment volatility is a separate, measurable source of noise

The offline snapshot mode isolates what happens when the same model runs the same benchmark in a stable environment versus a volatile one. Environmental drift, API changes, and service variation contribute to score variance that is unrelated to model capability. The framework makes this variance measurable for the first time across a common set of benchmarks.

For teams running internal evaluations against live environments, this finding suggests their results carry a noise floor they may not have accounted for.

Finding 3: The framework enables failure attribution that single-benchmark evals cannot

By standardizing resource reporting and failure logging across seven benchmarks and fifteen models, the framework produces the first cross-benchmark failure attribution analysis at this scale. Teams can identify whether a model's failures concentrate in specific task types, tool use patterns, or multi-step coordination scenarios. This is difficult to do when each benchmark produces different logs in different formats.

Scope and limitations

The framework standardizes the harness but cannot eliminate all evaluation confounds. The benchmark tasks themselves were designed by different teams with different goals. A unified harness reduces one source of noise; it does not eliminate differences in task distribution, difficulty calibration, or annotation quality between benchmarks. The paper is transparent that the goal is disentanglement, not elimination of all variance.

What this means
for choosing and
evaluating models.

The core implication is not complicated: before trusting a published leaderboard to inform a model selection decision, ask what harness produced it. Then run the candidates through the harness you will actually use.

1
For teams evaluating models for a specific task
Do not use published benchmark rankings as a proxy for performance in your setup. Build an internal evaluation suite that uses your own tool definitions, prompt format, and environment. Run candidate models through that suite. The model that wins on someone else's harness may not be the model that performs best in yours.
2
For teams running internal agent evaluations
Measure and report your harness configuration alongside your scores. If your evaluation results are going to inform a decision or be shared with stakeholders, the prompt format, loop architecture, tool schema, and environment stability are all part of what produced the number. Results from a different harness are not comparable, even on the same benchmark tasks.
3
For teams using live environments in evaluation
Quantify environment volatility before drawing conclusions from your evaluation scores. Run the same model multiple times under identical task specifications and measure how much variance comes from environment drift versus model behavior. If the variance is meaningful, switch to snapshotted environments for evaluation and reserve live testing for final validation.
4
For researchers publishing benchmark results
Report harness configuration as a first-class element of your evaluation, not a footnote. The instruction format, loop structure, and environment setup are as relevant to reproducibility as the dataset. Without them, other teams cannot determine whether a score difference is a model effect or a scaffold effect.
5
For business leaders using vendor benchmark claims
Ask which harness the benchmark was run with and whether any cross-harness comparison was performed. A model that was developed and evaluated on a tightly coupled scaffold has a higher probability of looking good on its home benchmark and a lower probability of generalizing. The useful comparison is cross-harness, not same-harness.

Where to go
from here.

Next steps if this work is relevant to how you evaluate or select models.

1
Read the paper
Zhu, P., Li, L., Lyu, Y., Luo, Q., Yang, J., Liu, Y., Hui, T., Yuan, X., Sun, L., Su, S., & Shao, J. (2026). A Unified Framework for the Evaluation of LLM Agentic Capabilities. Beijing University of Posts and Telecommunications, Shanghai AI Lab & collaborators. arXiv:2605.27898.
2
Audit your current evaluation harness
Write down the prompt format you use, the loop structure (ReAct, chain-of-thought, direct), how you define tool schemas, and whether your environment resets between runs. Then ask whether any benchmark scores you currently use were produced with a different setup. If so, they are not directly comparable to your internal results.
3
Test environment volatility on your current eval setup
Run your best-performing model on your evaluation tasks three times over three days without changing any model or prompt configuration. If the scores vary by more than a few percentage points, your environment is contributing noise you should account for before using those scores in a decision.
4
Read the related harness-effect work
This paper sits alongside two others in the Tandemly queue that explore how the harness layer shapes agent outcomes: Meta-Harness (Stanford and MIT, showing a 6x performance spread from harness variation alone) and Is Grep All You Need? (showing that harness choice shifts accuracy more than retrieval method across four agent frameworks). Together they build the case that harness is a first-class accuracy variable.
5
Build cross-harness evaluation before scaling model comparisons
When you need to compare two or more models for a production decision, run each model through at least two different harness configurations on the same task set. A model that consistently outperforms across harnesses is a stronger candidate than one that looks good only under the configuration it was promoted with.