Contract Net Protocol: Multi-Agent Task Auction Simulator (2D)

Auction Parameters

Worker agents 6

Task arrival rate 1.0/s

Bid noise (honesty) 15%

Load penalty weight 1.0

Live Metrics

Tasks awarded

Avg bids/task

0.0

Avg winning cost

0.0

Queue depth

Random-baseline cost

0.0

Auction efficiency gain

Protocol Log

How it works

The Contract Net Protocol (Smith, 1980) is a market-based mechanism for allocating tasks among autonomous agents without central control over which agent does the work — only the manager role is fixed:

1. announce(task)         manager broadcasts a task to all workers
2. bid_i = c_i(task) + w·load_i + ε_i   each worker estimates its own cost
3. award = argmin_i bid_i               manager awards to the lowest bidder
4. worker updates load_i += 1, decays over time

Here c_i(task) is the 2D Euclidean distance from worker i to the task marker, w is the load penalty weight, load_i is the worker's current queue, and ε_i is Gaussian-like bidding noise (the bid noise slider) representing imperfect self-estimation — a real LLM agent estimating its own capacity is never perfectly accurate either.

Worker agents — how many bidder nodes are in the network; more agents means more competition per task.
Task arrival rate — how often the manager announces a new task.
Bid noise — variance added to each bid; at 0% the protocol always awards the true lowest-cost agent, at high values awards become closer to random.
Load penalty weight — how strongly a busy agent's own backlog raises its bid, which is what keeps the protocol from swamping a single fast agent with every task.

Each awarded task is also scored against a random-assignment baseline — the true distance cost had the manager simply handed the task to a random worker instead of auctioning it — so the running "auction efficiency gain" stat shows the real, measured savings the auction mechanism produces over no coordination at all.

Real-world relevance: this exact announce→bid→award loop is how many multi-agent LLM orchestration frameworks route a task to the best-suited tool-using agent, and it underlies distributed job schedulers, robotic task allocation, and decentralized marketplaces.

Manager Idle worker Bidding Awarded task