Skip to content

Multi-Agent Swarm

AI agents that check their own work

Parallel agents split your work into parts and do them at the same time. Then code, not the model's opinion, checks every result: quotes against the pages they came from, apps against their own tests, numbers against your data. Free and open source; bring any model.

A real run, replayed: swarm build writes a FastAPI + SQLite app in six parallel waves, then installs it and runs its 22 tests. Total cost on DeepSeek: $0.0057.

Try it in a minute, free

pip install "multi-agent-swarm[all]"
swarm research "How do heat pumps compare to gas boilers in cold climates?" --mock --out demo.html

--mock runs the whole pipeline offline, with no keys and no cost, so you can see a cited report before you spend anything. For real runs, add a model key and a free search key:

export ANTHROPIC_API_KEY=...     # or any OpenAI-compatible provider: see Configuration
export TINYFISH_API_KEY=...      # free web search: https://agent.tinyfish.ai/api-keys
swarm research "How do heat pumps compare to gas boilers in cold climates?" --out report.html

What it does, and how each result is checked

Job Command Checked by
Research a question, with citations swarm research every quote matched word for word to the page read; each cited claim checked against its quotes
Build an app (backend, frontend, CLI) swarm build "…" --out app --allow-execution the app's own tests, installed and run; failures go to a fixer
Analyze a CSV swarm analyze data.csv "…" --allow-code-execution every number must come from code that ran on your data
Review a codebase swarm review ./repo each finding must quote real code; a second check against the imported files refutes the ones that don't hold
Write a long document swarm write "…" --words 3000 an outline every section follows; an edit pass that must keep every section
Translate, summarize or rewrite long documents swarm doc translate file.md --to German one glossary for every chunk; length checks
Thousands of small tasks swarm run tasks.jsonl exact answers, patterns, sandboxed tests, or grounded sources

Real results

Live runs on DeepSeek V4.1 Flash, with web search free through TinyFish:

Run Result Cost
Build a FastAPI + SQLite todo API 14 files, 22 passing tests, 0 fix rounds $0.0057
Research "heat pumps vs gas boilers in cold climates" 3 of 3 sub-questions verified, 19 sources, 4 unmatched quotes dropped $0.016
Analyze a 144-row sales CSV every figure matched the raw data $0.0024
Review an 11-file codebase 6 findings kept; 4 refuted on a second check, listed with reasons $0.0030

Every run has a hard cap (--max-usd), and the estimate comes before any spend.

Where to go next