The Four Agents

Four AI components cover the whole market lifecycle - three of them autonomous, plus a co-pilot that serves users on request. The claim is specific: the AI trades, creates, makes, and settles the markets - one agent per verb.

Co-Pilot - trades

Parses natural language into structured trade intents, quotes odds and payouts, drafts new markets on request, and remembers context across sessions per wallet. Full detail in The AI Co-Pilot.

Pulse Drafter - creates

Every morning the drafter asks the intent engine to write the day's machine-resolvable markets and deploys them onchain through the same permissionless createMarket anyone else uses. Its default repertoire:

• PULSE - X Layer's own public metrics: daily active wallets, OKB 24h trading volume (all venues, per CoinGecko), X Layer total DeFi TVL and stablecoin supply (per DefiLlama)
• RWA - total tokenized real-world-asset TVL across DeFi (per DefiLlama)
• EQUITY - whether TSLA / NVDA close above a specific price at that day's official Nasdaq close - skipped automatically on weekends and US market holidays, because a market naming a non-trading session could never settle

Drafts must name the metric, an "above " comparator, the window, and the public data source - the wording discipline that makes autonomous settlement possible. Drafts whose close time falls outside a 2-48h window are skipped, never force-clamped: rewriting "by Sep 12" into a 48-hour market would settle the wrong question.

Market Maker - makes

A continuous loop that keeps markets tradeable:

• Bootstraps empty markets two-sided, split by the AI's fair-odds estimate (from /api/ai/odds) so fresh markets open near fair implied probability instead of a meaningless 50/50
• Tops up thin sides so takers always have a counterparty
• Operates under hard caps: lifetime budget per run, per-stake maximum, a gas reserve floor, and affordability checks before the paid AI call - it never spends inference on a bootstrap it cannot fund
• Its stakes are real parimutuel positions - when a market resolves against it, its stake pays the winners. Depth provision is inventory risk, not wash activity - and it never touches OKX DEX by design

Resolver - settles

The agent most venues get wrong, so it gets its own page: Resolution & Settlement. The one-line version: it settles exactly what is mechanically checkable from public data, and refuses everything else to a human review queue - because resolveMarket is final.

Watching them work: the AGENT OPS console

Every agent posts each decision to /api/agent-log - liquidity seeded (and at what AI ratio), markets drafted, settlements taken or refused, heartbeat scans. The AGENT OPS console renders this as a live terminal with per-agent status.

The roster is honest about cadence. Cron agents show ONLINE for the minutes around a pass, then LAST 52M AGO - elapsed time since they last worked. A resolver showing "LAST 52M AGO" at 07:53 is not a fault: its last pass was the 07:00 cron, exactly on schedule, and it will flip back online at 08:00. Watching the timestamps track the cron schedule is the proof the fleet is real - four permanently green dots would be the suspicious display.

Reports are authenticated (AGENT_LOG_SECRET) and observability-only: an unreachable console never blocks trading, drafting, or settlement. Every stake, deploy, and settlement in the feed is a real onchain action you can cross-check in the activity ticker or on the X Layer explorer.

AI infrastructure

Inference runs on a provider abstraction: 0G Compute (decentralized, OpenAI-compatible - the live default) or Anthropic (schema-guaranteed structured outputs). Open-model outputs are treated as untrusted input: schema-validated, coerced, and guarded server-side before anything reaches an execution path.