Open Agent Leaderboard
Collection
An open benchmark for comparing full agent systems across diverse real-world tasks. Reports both quality and cost. • 4 items • Updated • 1
Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html>
<h"... is not valid JSON
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Open Agent Leaderboard Results
Detailed evaluation results for general-purpose AI agents across diverse real-world benchmarks — without domain-specific tuning.
- Leaderboard: open-agent-leaderboard/leaderboard
- Website: exgentic.ai
- Paper: arXiv:2602.22953
- GitHub: Exgentic/exgentic
- License: CDLA-Permissive-2.0
Benchmarks
| Benchmark | Task ID | Description |
|---|---|---|
| AppWorld | appworld |
App-based task completion in simulated smartphone environments |
| BrowseComp+ | browsecomp_plus |
Web browsing and complex information retrieval |
| SWE-bench | swebench |
Software engineering issue resolution on real GitHub repos |
| TauBench-Airline | taubench_airline |
Customer service agent evaluation (airline domain) |
| TauBench-Retail | taubench_retail |
Customer service agent evaluation (retail domain) |
| TauBench-Telecom | taubench_telecom |
Customer service agent evaluation (telecom domain) |
The overall score is a weighted average: each TauBench sub-task gets 1/12 weight (1/4 total for TauBench), all others get 1/4 each.
Agents Evaluated
| Agent | Framework |
|---|---|
| Claude Code | claude-code |
| OpenAI Solo | openai-agents-python |
| Smolagent | smolagents |
| React | litellm |
| React + Shortlisting | litellm + exgentic |
Models
Results are reported for each agent × model combination: Claude Opus 4.5, Gemini 3 Pro, GPT-5.2, DeepSeek V3.2, Kimi K2.5.
Submitting new results
This dataset is the source of truth for the Open Agent Leaderboard. To add results for a new model, agent, or benchmark:
- Run evaluations using the Exgentic framework
- Open a PR on this dataset adding your rows to the parquet file in
data/
Each row represents one (agent, model, benchmark) combination. Required fields:
| Field | Description |
|---|---|
agent |
Agent identifier (e.g., claude_code) |
agent_name |
Display name (e.g., Claude Code CLI) |
model |
Model identifier (e.g., openai_Azure_DeepSeek-V3.2) |
model_name |
Display name (e.g., openai/azure/DeepSeek-V3.2) |
benchmark |
Benchmark identifier (e.g., swebench) |
benchmark_name |
Display name (e.g., SWE-bench) |
benchmark_score |
Primary score (0-1) |
planned_sessions |
Number of tasks attempted |
total_sessions |
Number of sessions completed |
successful_sessions |
Number of sessions that passed |
See the existing data for the full schema and examples.
Schema
| Column | Type | Description |
|---|---|---|
agent / agent_name |
string | Agent identifier and display name |
model / model_name |
string | Model identifier and display name |
benchmark / benchmark_name |
string | Benchmark identifier and display name |
benchmark_score |
float | Primary success rate (0-1) |
average_score |
float | Average score across sessions |
average_agent_cost |
float | Average cost per task (USD) |
average_steps |
float | Average number of agent steps per task |
average_action_count |
float | Average number of actions per task |
average_invalid_action_count |
float | Average invalid actions per task |
percent_successful |
float | Fraction of tasks that succeeded |
percent_finished |
float | Fraction of tasks that completed (success or fail) |
percent_error |
float | Fraction of tasks that errored |
total_agent_cost |
float | Total cost across all tasks (USD) |
planned_sessions / total_sessions / successful_sessions |
int | Session counts |
- Downloads last month
- 82