AuctionSI

Let agents compete to solve the work.

AuctionSI is an open-source marketplace and auction framework for autonomous agents. Instead of a central orchestrator choosing which agent does a task, the task is announced to a market, capable agents bid, an auction mechanism picks the winner, and the work is contracted, executed, verified, settled and fed back into reputation.

Task → discovery → bids → auction → winner → contract → execution
     → verification → settlement → reputation

Agents can be Python functions, HTTP services, solvers, ML models, humans, LLMs or simulations. Nothing requires an LLM, an API key or a cloud service. Everything runs locally, and simulations are deterministic when seeded.

The auction mechanism, the winner-selection policy, the bidding strategy, the verifier, the settlement policy, the reputation system and the experiment are all plugins.

Install

pip install auctionsi           # Python 3.11+
pip install -e ".[dev]"        # from a clone, for development
pip install "auctionsi[plotly]"  # optional figures

Quick start

from auctionsi import Marketplace, PythonFunctionAgent, RecoveryPolicy, Task
from auctionsi.verification import ExactMatchVerifier

def right(task, contract):
    return {"answer": 42}

def wrong(task, contract):
    return {"answer": 41}

market = Marketplace(
    verifier=ExactMatchVerifier(42, key="answer"),
    recovery=RecoveryPolicy(next_best=1),
)
market.register(PythonFunctionAgent("agent-a", right, capabilities=["qa"], price=0.05))
market.register(PythonFunctionAgent("agent-b", wrong, capabilities=["qa"], price=0.07))
market.register(PythonFunctionAgent("agent-c", wrong, capabilities=["qa"], price=0.03))

result = market.submit_task(Task(task_id="task-001", task_type="qa", budget=0.10))
print(result.explain())                 # ranking with per-term score contributions
print("\n".join(result.trace()))        # every lifecycle event
assert result.winner == "agent-a"       # C was cheapest, failed verification, A took over

What is in v0.1

Not in v0.1 (see the roadmap): REST/WebSocket API, web dashboard, distributed execution, combinatorial and forward auctions, learning bidders.

Simulate and experiment

from auctionsi.experiments import compare_mechanisms
from auctionsi.mechanisms import FirstPriceReverseAuction, SecondPriceReverseAuction

result = compare_mechanisms(
    [FirstPriceReverseAuction(), SecondPriceReverseAuction()],
    environment={"agents": {"count": 20}, "tasks": {"count": 100}},
    replications=5,
    seed=42,
    metrics=["total_cost", "completion_rate"],
)
for c in result.comparisons():
    print(c.metric, c.treatment, round(c.mean_diff, 4), (round(c.ci_low, 4), round(c.ci_high, 4)))

auctionsi market simulate --agents 100 --tasks 1000 --seed 42
auctionsi experiment run research/experiments/mechanisms.yaml

Six reproducible experiments (mechanisms, market size, reliability, reputation, strategies, concentration) with their reports are in research/. All results there come from synthetic markets and say so.

Examples

examples/: data-analysis (fail-over after deterministic checks), SQL (queries executed against a read-only test database), scheduling optimisation (constraint and objective checks), research reports (structure and citation checks), compute capacity (constraints and multi-winner uniform pricing), and a simulated collusion experiment. python scripts/run_examples.py runs them all.

Benchmarks

Measured on an Apple-silicon laptop, Python 3.12 (details and method in benchmarks/):

Documentation

docs/: concepts, architecture, mechanisms (with assumptions and limitations), selection, simulation and metrics, extending, CLI, and security.

Roadmap

• v0.1 core marketplace (this release)
• v0.2 more mechanisms (scoring auctions, reserve-price variants)
• v0.3 richer simulation and experiment designs (factorial sweeps)
• v0.4 reputation research and stronger adversarial models and defences
• v0.5 persistence backends (PostgreSQL) and distributed execution
• v0.6 REST/WebSocket API and web dashboard
• v0.7 agent protocol interoperability
• v0.8 advanced mechanism research (combinatorial, forward, capacity auctions, learning bidders)
• v1.0 stable public API

No dates are promised.

Contributing and licence

See CONTRIBUTING.md and CODE_OF_CONDUCT.md. MIT licensed (LICENSE, NOTICE).