Minimal agent harness that <120 lines for web-search tasks. Trust the model, minimize the wrapper.
Model: Minimax API (function calling)
Search: minimax-coding-plan-mcp MCP server (same API key, spawned once per run)
Tools: web_search, bash
Loop: append-only message history; stops when the model returns no tool calls
uv sync
cp .env.example .env # fill in MINIMAX_API_KEY
python agent.py "Who founded Shiseido and in what year?"python -m bench --tasks data/gaia_val.csv --out results/gaia.csv --preset gaiaDownloads 200 tasks from HuggingFace automatically. The agent searches the web and returns a Markdown table; evaluation scores each cell with metrics like exact match, number proximity, URL domain match, date proximity, and LLM-as-judge.
# Download dataset
python -m bench.widesearch download
# Run 10 English tasks
python -m bench --out results/ws.jsonl --preset widesearch --lang en --limit 10
# Run all English tasks
python -m bench --out results/ws.jsonl --preset widesearch --lang en
# Resume an interrupted run (completed tasks are skipped)
python -m bench --out results/ws.jsonl --preset widesearch --lang enpython -m bench --tasks data/custom.csv --out results/custom.csv \
--system "You are a precise assistant. Answer concisely."Enrich and score companies from a CSV against user-defined criteria.
# With context file (criteria, instructions, examples)
python -m bench --preset enrich --tasks data/companies.csv \
--out results/leads.csv --context data/criteria.md
# With inline criteria
python -m bench --preset enrich --tasks data/companies.csv \
--out results/leads.csv --criteria "B2B cosmetics brands with DTC in Japan"Outputs two files:
results/leads.csv— all rows, enriched with score/qualifies/reasoningresults/leads_qualified.csv— only qualifying rows
See data/test_criteria.md for context file format.
| Preset | Use for |
|---|---|
gaia |
GAIA Level 1-3, concise single-answer questions |
widesearch |
WideSearch, structured table collection from live web |
enrichment |
Japanese company name resolution, JSON output |
enrich |
Company enrichment & scoring against user-defined criteria |