KDD Cup 2026 Data Agents

Mamba Agent
Mamba Never Out — in memory of basketball legend Kobe Bryant.

ReAct-style data agent for the DABench benchmark. The runner reads tasks from data/public/input/, drives an OpenAI-compatible LLM through a fixed tool inventory, and writes per-task prediction.csv files plus a per-run summary.json. Optional SQLite tracing writes debug spans to artifacts/traces.db for the dashboard.

Quick Start

1. Install uv — on macOS / Linux:

   curl -LsSf https://astral.sh/uv/install.sh | sh

2. Sync dependencies:

   uv sync

3. Confirm the dataset is visible:

   uv run dabench status --config configs/react.example.yaml

4. Run the agent:

   uv run dabench run-benchmark --config configs/react.example.yaml

For more, see docs/: architecture.md, cli.md, agent-protocol.md, development.md.

Dataset

The public demo dataset lives under data/public/input/. Each task directory follows:

data/public/input/task_<id>/
├── task.json
└── context/

Public answers live separately under data/public/output/task_<id>/gold.csv. Hidden test sets only ship input/.

task.json carries task_id, difficulty, and question. context/ may contain CSV, JSON, SQLite/DB, and text files.

Configuration

The YAML schema is summarized in docs/cli.md, with validation implemented on the frozen config dataclasses in src/agents/config.py. An example lives at configs/react.example.yaml.

dataset:
  root_path: data/public/input

agent:
  model: YOUR_MODEL_NAME
  api_base: YOUR_API_BASE_URL
  api_key: YOUR_API_KEY
  max_steps: 16
  temperature: 0.0

run:
  output_dir: artifacts/runs
  run_id:
  max_workers: 4
  task_timeout_seconds: 600

CLI

uv run dabench <command> --config PATH [options]

Command	Purpose
status	Show project paths, config path, dataset root, and public task counts.
inspect-task	Show task metadata and list accessible files under context/.
run-task	Run the agent on one task.
run-benchmark	Run the agent across the dataset. Accepts --limit N and --range S-E.
dashboard	Launch the SQLite traces dashboard. Requires runs created with tracing.enabled: true.

Tools

The agent calls into the default tool inventory (registered in src/agents/tools/registry.py):

Tool	Inputs
inspect_files	(none)
preview_file	path
execute_context_sql	path, sql, limit
execute_python	code
answer	columns, rows, from_csv

All path arguments are relative to the task context/ directory.

Outputs

artifacts/runs/<run_id>/
├── summary.json
└── <task_id>/
    └── prediction.csv

summary.json is the canonical machine-readable run output. When a config contains tracing.enabled: true, tracing data is written to tracing.db_path (artifacts/traces.db by default) and can be viewed with uv run dabench dashboard.

Main Modules

Module	Responsibility
src/agents/benchmark/dataset.py	Public dataset loader
src/agents/tools/{filesystem,sqlite,python_exec,registry}.py	Tool inventory + sandboxed handlers
src/agents/agent.py	ReAct loop driver
src/agents/runtime/{state,messages,media,budget,recorder}.py	Run state, message replay, video attachment, budget guardrails, step records, and callbacks
src/agents/llm/{openai,protocol,rate_limit,token_bucket}.py	OpenAI adapters, response protocol adapter, and rate limiter
src/agents/verification/{answer,terminal_policy}.py	Optional answer self-check and terminal-answer fallback policy
src/agents/application.py + src/agents/runs/	App factory + subprocess runner
src/agents/dashboard/ + dashboard/frontend/	Traces dashboard backend and Vite frontend source
src/agents/cli.py	dabench entrypoint