AGENT / HARNESS

From one attempt to an execution loop.

The model proposes the analysis. The harness manages tools, feedback, budgets and final artifacts. InferenceNet makes the resulting process and output inspectable.

01 / EXECUTION MODES

The harness is part of the experiment.

A / BASELINE

Single-pass Agent

1 model call

Generate one program, then execute it. There is no iterative tool-feedback loop.

See published results →

B / AGENT LOOP

DeepAgents

≤ 6 model calls

Plan, inspect data, execute trial programs and revise. Up to four trial tool calls, followed by final execution.

Read the full walkthrough ↗

C / INDEPENDENT

DeepSeek Harness · DSH

≤ 6 calls / attempt

Official DSH SDK. The published run uses technical recovery and stays separate from the baseline / DeepAgents paired comparison.

Read the run protocol ↗

These are the published run settings, not universal framework limits. Unequal total budgets and differing protocols do not support an equal-cost causal claim.

02 / FEEDBACK

A loop with an explicit stopping point.

  1. 01

    Read the task

    A research specification and its source data.

  2. 02

    Write & run code

    Fit the requested econometric model.

  3. 03

    Inspect & revise

    A harness can use execution feedback.

  4. 04

    Recover the finding

    Return the coefficient, standard error and p-value.

↶ Execution feedback returns to the model until it finalizes or reaches its budget.

03 / RECORDED CASE

Same task. Two execution paths.

Task 0011 asks for a monthly-clustered regression on Spain, 1905–1945. This recorded GPT-5.5 example shows why inspecting the data can change an outcome.

SINGLE-PASS

A guessed schema stops execution.

The generated program searches for common month/date column names. It misses the dataset’s tid column and exits before producing a result.

1 model call · no valid result JSON

DEEPAGENTS

Inspect, estimate, then finalize.

The agent inspects 38 columns, checks the time structure, then fits OLS with tid-clustered errors on 977 observations. Its final program recovers all three statistics.

4 model calls · 3 tool calls

The final artifact

A program writes these three numeric fields. Execution logs and the evaluation record are stored separately.

{
  "coefficient": 0.27273630753072026,
  "standard_error": 0.054224701707830926,
  "p_value": 6.899021793750105e-7
}

Illustrative internal run from 13 September 2026, not a result in the current leaderboard. One example does not establish an average harness benefit.

Inspect the complete trace ↗ Scoring definitions & evidence →