Hugging Face's logo

Datasets:
Exgentic
/
agent-cards

Dataset card Files
xet
 Community
agent-cards
File size: 2,841 Bytes
1a17759
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
---
agent_name: "React + Shortlisting"
version: "exgentic 0.1.0 / litellm 1.79.1"
developers: "Exgentic Team"
license: "Apache-2.0"
framework: "Exgentic"
repository: "https://github.com/Exgentic/exgentic"
models:
  - "Claude Opus 4.5"
  - "GPT-5.2"
  - "Gemini Pro 3"
tags:
  - agent
  - react
  - tool-calling
  - general-purpose
---

# React + Shortlisting

> A general-purpose ReAct agent with tool shortlisting -- helps the agent focus on relevant tools instead of searching through everything.

## Agent Details

- **Name**: React + Shortlisting
- **Version**: exgentic 0.1.0 / litellm 1.79.1
- **Developers**: Exgentic Team
- **Framework**: [Exgentic](https://github.com/Exgentic/exgentic)
- **License**: Apache-2.0

## Architecture

### Planning
ReAct (Reasoning + Acting) loop: the agent reasons about the current state, selects an action, observes the result, and repeats.

### Memory
Full conversation history within the session. No cross-session memory.

### Tool Use
Function calling via LiteLLM. Before each step, a shortlisting phase narrows the available tools to the most relevant subset for the current task context. This reduces noise and improves action selection, especially in benchmarks with large action spaces.

### Error Recovery
Retry on transient failures. The agent can observe error messages and adjust its approach on subsequent steps.

### Subagents
None -- single-agent system.

## Models

| Role | Model | Purpose |
|------|-------|---------|
| Main | Claude Opus 4.5 / GPT-5.2 / Gemini Pro 3 | Reasoning, action selection, tool calls |

## Supported Environments

- AppWorld (app-based task completion)
- BrowseComp+ (web research)
- SWE-bench (software engineering)
- TauBench Airline, Retail, Telecom (customer service)

## Evaluation Results

| Benchmark | Score (Opus 4.5) | Score (Gemini Pro 3) | Avg Cost (Opus 4.5) |
|-----------|-------------------|----------------------|----------------------|
| AppWorld | 0.66 | 0.40 | $7.80 |
| SWE-bench | 0.49 | 0.36 | $3.64 |
| TauBench-Airline | 0.56 | 0.59 | $0.96 |
| TauBench-Retail | 0.72 | 0.72 | $1.06 |

**Full results**: [Open Agent Leaderboard](https://huggingface.co/spaces/open-agent-leaderboard/leaderboard)

## Cost Profile

The most cost-efficient configuration in the top 5 on the leaderboard. With Gemini Pro 3, average cost per task is $0.66 -- a fraction of other top agents. Tool shortlisting reduces unnecessary tool calls, directly lowering cost.

## Limitations

- No cross-session memory or learning
- Performance depends heavily on the quality of tool descriptions
- Shortlisting may occasionally filter out a relevant tool in novel contexts

## How to Run

```bash
uv tool install exgentic

exgentic evaluate \
  --benchmark tau2 \
  --subset retail \
  --agent tool_calling_with_shortlisting \
  --model openai/claude-opus-4-5
```