stampede

The wind tunnel for the agent economy

You built it for agents. stampede shows you what agents will actually do to it. Generate realistic (and adversarial) agent populations, turn them loose on your MCP server, API, or protocol in a sandbox, and get a report of where they succeed, where they get confused, and where they break you.

▶ demo GIF coming — the watchable swarm crashing & recovering against a live target

Status: 🚧 v0.1 in progress. Building in public — follow the "wind tunnel" thread series.

---

Why

Software is increasingly consumed by agents, not humans. But if you build an MCP server, tool API, or onchain protocol, you have no way to know how agents behave against it before shipping. You test one agent on one happy path, or fire uniform load that doesn't speak agent protocols. Both miss the real risk: agents don't generate traffic, they generate behavior — they misread tool descriptions, pick the wrong tool, retry in loops, and (when adversarial) exploit.

stampede generates a population of heterogeneous, stateful, realistically-flawed agents and turns them loose on your system to see what breaks.

Quickstart

# from source (v0.1 — not yet on PyPI):
git clone https://github.com/swarmproof/stampede && cd stampede
python -m venv .venv && .venv/bin/pip install -e .

stampede init                       # writes a starter stampede.yaml (targets a mock world)
stampede run --dry-run              # zero-LLM, deterministic — watch the misuse map appear
# report at ./stampede-report.html

# against your own MCP server, and live (needs the extras + a model):
stampede run --target "python my_server.py" --size 200 --live

--dry-run needs no API keys and no network: it runs the deterministic heuristic swarm so you see the report shape in seconds. Drop it (and set real models:) for a live run. See examples/ for MCP / HTTP / mock walkthroughs.

The signature artifacts

• Persona packs — versioned YAML defining agent temperaments, not just models: naive, expert, impatient, frugal, adversarial, drunk. Community-contributable.
• The Agent Readiness Report — the screenshotable deliverable: task-success by persona, the misuse map ("34% of agents called delete_record when they meant archive_record"), swarm concurrency (p95/p99), the cost profile, and adversarial findings.
• The watchable world — a live local dashboard where you see the swarm hit your system and click any agent to ask "why did you call that tool?"
• The Agent Ready badge — drop --badge agent-ready.svg (or --summary summary.json) on any run and put the grade in your README, the way web teams post a Lighthouse score. Pair it with stampede diff for a CI gate that flags real regressions, not RNG noise.

See SPEC.md for the full design and ROADMAP.md for what ships when.

Part of the Swarm Proof toolkit

Trust infrastructure for the agent economy — seven projects, one thesis.

Project	What it does
stampede ← you are here	Point a herd of realistic agents at your system before real ones arrive
mockworld	A synthetic internet for agents — fake Stripe, Gmail, exchange, instantly
mcp-probe	The CI quality suite for MCP servers — lint, contract-test, benchmark, load
costbomb	Denial-of-wallet fuzzing — find the inputs that make your agent spend $500
exactly-once	Idempotency middleware so agent side-effects fire once
agent-postmortems	A structured incident database + post-mortem standard for agent failures
awesome-agent-reliability	The curated map of the field

License

Apache-2.0. Provider-agnostic (OpenAI-compatible + Anthropic SDK + Ollama). Citable via CITATION.cff.