Skip to content

Roadmap

No dates are promised. "Proposed" means not started; anything not listed under a release does not exist yet.

Released

0.1 - the core marketplace: tasks, agents, validated bids, auction and contract state machines; first-price, second-price, multi-winner and open descending reverse auctions; explainable selection policies; deterministic verifiers; settlement; multi-dimensional reputation with decay; recovery (retry, fail-over, re-open); events, traces and deterministic replay; SQLite; simulation, metrics and replicated experiments; Python, HTTP, OpenAI-compatible and human agents; the CLI.

0.2 - forward, bundle (combinatorial) and capacity auctions; a learning (bandit) bidder; exploration-bonus selection; adversaries, mid-run changes and factorial sweeps in experiment configs; calibration reports; HMAC-signed bids and bid and execution timeouts; PostgreSQL; the REST/WebSocket API and dashboard.

Unreleased - agent removal across API, CLI and dashboard; WebSocket authentication without URLs; dashboard reconnects and tests; see the changelog.

Proposed

Mechanisms and selection - Scoring auctions (announced scoring rules over price and quality) and reserve-price variants for reverse auctions. - Larger bundle auctions. The exact dynamic program is exponential in the number of items, so BundleReverseAuction caps it (max_items, default 12); larger instances need a solver or heuristics. - Double auctions, where several buyers and sellers trade at once.

Market integrity - Per-auction and per-group collusion screens. The collusion experiment found that the market-wide averages of the current screens did not detect a simple bid-rotation cartel. - Identity verification behind metadata["operator"], which the Sybil cap relies on.

Operations - Distributed execution: agents and auctions across processes or machines (today one process runs the market, with concurrent bid solicitation in a thread pool). - Multi-tenant API with per-user keys and permissions (today one shared key). - Sandboxing for in-process Python agents.

Interoperability - Adapters for standard agent-to-agent protocols, so agents built for other frameworks can bid without custom code.

Stability - 1.0: a stable public API with a deprecation policy.

Known limitations

  • Incentive properties (for example truthful bidding under second price) hold only under the assumptions stated in auction mechanisms; learning bidders in the research results do not bid truthfully.
  • Research results come from synthetic markets, not measurements of real agents.
  • The SQLite store expects one writing process.

Suggestions and contributions are welcome: see contributing.