The three tiers of prediction market automation
"Automated betting tools" and "prediction market bots" cover three structurally different things. Knowing which tier you are looking at is the difference between a useful tool and a money pit.
| Tier | What it does | Decides trades? | Examples |
|---|---|---|---|
| 1. Signal tools | Surface data, odds, mispricings: you act | No | Dashboards, odds scanners, alert bots |
| 2. Execution bots | Place/cancel orders on hardcoded rules | No (follows rules) | py-clob-client scripts, arbitrage bots |
| 3. AI agents | Read context, decide, execute autonomously | Yes | NickAI, self-hosted multi-LLM agents |
Tier 1: Signal tools
The largest and most crowded tier. Signal tools surface information: live odds across venues, order-book depth, historical resolution rates, cross-venue spreads. They do not trade; they inform a human who trades.
What works. Cross-venue odds scanners (spot Polymarket vs Kalshi divergence), order-book depth monitors, and resolution-history trackers. These are genuinely useful inputs.
What to skip. "AI-powered signal" tools that are rule-based engines with an LLM label. The signal is the data; the "AI" is usually marketing. Pay for the data feed, not the wrapper.
Tier 2: Execution bots
Rule-following automation. You define the logic, "buy YES if the spread between Polymarket and Kalshi exceeds 2%", and the bot executes it deterministically. The workhorses here are py-clob-client scripts (Polymarket) and Kalshi API scripts.
What works. Arbitrage bots (cross-venue, YES+NO mispricing, multi-outcome sum≠100%), market-making bots quoting both sides, scheduled execution. Deterministic strategies where the rules are stable.
What to skip. Execution bots that hold your funds. Any tool asking for a deposit to a platform address is custodial: an unregistered exchange with the multi-cycle failure history that implies. Non-custodial (your wallet, your keys) is the only safe execution model.
Tier 3: AI agents
The only tier where software actually decides. An AI agent reads news, event data, and market context, estimates a fair probability via LLM reasoning (ideally multi-model consensus), and trades when the market is mispriced: adapting to conditions the original config did not anticipate.
What works. Multi-LLM consensus agents that trade non-custodially through your own wallet or API key, with hardcoded policy caps and per-trade audit trails. This is the tier that captures news-reactive and niche-event edge a rule-based bot cannot.
What to skip. Single-LLM agents (calibration and hallucination risk), and any "AI agent" that is actually a Tier 1 signal tool or Tier 2 rule bot with an AI label. The test: does it decide and execute on its own, reading real-time context? If not, it is Tier 1 or 2.
The custody dividing line runs through every tier
Regardless of tier, the single most important property is custody. Two questions before using any prediction-market automation tool:
- Where do my funds sit? Your wallet / your exchange account (non-custodial) or the tool's address (custodial)?
- What permissions does it need? Wallet-signature or trade-only API key (safe) vs deposit or withdrawal permission (unsafe)?
Non-custodial tools can lose you trading PnL; custodial tools can lose you everything. See non-custodial vs custodial AI bots for the full failure-mode taxonomy.
Which tier do you actually need?
- You want to trade manually with better information. Tier 1 signal tools.
- You have a fixed, rule-based strategy (arbitrage, market making). Tier 2 execution bots, non-custodial.
- You want software that reads the world and decides. Tier 3 AI agents, multi-model, non-custodial.
Most serious prediction-market traders end up combining Tier 1 (for their own monitoring) with Tier 3 (for autonomous execution on the strategies that benefit from real-time judgment). Tier 2 remains right for the genuinely deterministic strategies, pure arbitrage, market making, where an LLM adds cost without adding edge.
Building vs buying
Tiers 1 and 2 are buildable in a weekend to a few weeks depending on scope. Tier 3 is 2–4 engineer-weeks for a working v1 plus an operational tail (model versions drift, APIs change). NickAI provides Tier 3 as a runtime: multi-LLM consensus over Polymarket and Kalshi, non-custodial, with policy caps and audit built in. The full agent architecture is documented in how the NickAI prediction-market agent works.