When AI Agents Trade Against Each Other: Who Makes the Market?

Imagine two AI agents watching the same election market. Both estimate that “Yes” is worth about $0.60. One posts a bid at $0.58 and an ask at $0.62. The other spots breaking information, decides $0.60 is stale and immediately buys the $0.62 offer.

The first agent appeared to be earning a four-cent spread. Instead, it may have sold just before the fair value moved higher. That is the basic tension in market making: you are paid for standing ready to trade, but the person taking your quote may know something you do not.

For the non-AI foundation, read Maker, Taker and Arbitrageur and The Order Book Is Not the Market. The mechanics and risks exist whether the decision maker is a person, conventional bot or AI agent.

Prediction markets make this especially intuitive because a share settles at $1 if an outcome occurs and $0 if it does not. But the mechanics are still those of an order book. On Polymarket, prices emerge from users trading through a central limit order book (CLOB), according to its current prices and order-book documentation. A displayed 60% probability is not a guaranteed execution price.

So when AI agents trade against one another, who makes the market? Whoever is willing to post tradable bids and asks—and manage the risks that arrive with the fills. That could be an individual, a firm or automated software. The public book does not reliably reveal which one.

What a market maker actually does

A market maker provides resting liquidity. It posts a price at which it will buy (the bid) and a price at which it will sell (the ask). The gap is the bid-ask spread.

If the best bid is $0.58 and best ask is $0.62, the midpoint is $0.60 and the quoted spread is $0.04. A trader who wants immediate execution crosses the spread: buying at the ask or selling at the bid. Polymarket notes that its market orders are technically limit orders priced to execute immediately against resting orders. Limit orders can also partially fill and leave a remainder in the book.

The market maker is not simply collecting four cents. It may fill only one side. It may wait minutes or hours before offsetting the position. The event probability may change. Fees and incentives may apply. A quote may become stale before a cancellation reaches the matching engine.

The spread is compensation for those uncertainties, not free income.

Inventory: the position left behind

Every fill changes the maker’s inventory. If traders repeatedly sell “Yes” into its bid, the maker accumulates “Yes” shares. It now has directional exposure: if the market’s probability falls, that inventory loses value before it can be sold.

Market makers therefore skew price or size. A maker holding too much “Yes” might lower its bid, post a more attractive ask or quote less size on the buy side. The goal is to make further accumulation less likely and reduction more likely.

Polymarket’s official market-making guide tells makers to incorporate inventory into every pricing and sizing decision, track exposure by market and outcome, skew quotes to reduce imbalance and deliberately manage remaining directional exposure. The classic Avellaneda–Stoikov limit-order-book model formalizes the same core idea: a dealer’s preferred quote changes with its inventory and risk tolerance.

An AI can help estimate fair value or adjust quotes, but it does not remove inventory risk. In fact, an agent optimized only for fill count may acquire the worst possible inventory quickly.

Adverse selection: why did they trade with you?

Suppose a fast trader buys every time your ask is about to become stale, but ignores it when your price is safe. You will show many fills and may still lose money.

This is adverse selection: order flow can be informative. The trader choosing to hit a quote may have fresher news, a better probability model, a faster connection or a price discrepancy on another venue. The foundational Glosten–Milgrom market-microstructure paper showed how informed trading can produce a positive bid-ask spread even for a risk-neutral market maker with zero expected profit.

In a prediction market, the risk can jump around scheduled catalysts—an economic release, court decision, debate or resolution update. It can also appear without warning when a credible source breaks news. A maker that does not understand the market’s resolution rules may quote a probability for the headline while informed traders price the actual contract.

Measure adverse selection after the fill. Compare the execution price with the midpoint or a defensible fair-value estimate after fixed horizons. If buys are routinely followed by price rises and sells by falls, the maker is being picked off. Raw spread capture before this markout can be misleading.

Latency turns a good quote into a bad one

A market-making loop must receive a book update, calculate fair value, apply risk checks, submit or cancel orders and confirm the result. Delay at any stage creates stale-quote risk.

Polymarket provides WebSocket channels for near-real-time order-book, trade and user-order updates. Its guidance recommends using real-time updates rather than polling, batching related orders and fetching open orders and recent trades after reconnecting. It also says orders cannot be edited in place; changing a quote requires canceling and replacing it.

That creates important race conditions. A quote can fill while its cancellation is in flight. A network reconnect can leave the local view out of sync. One order in a submitted batch may fail while others succeed. A robust agent reconciles venue state instead of assuming its last request happened exactly as intended.

Speed is relative. A 300-millisecond loop may be adequate in a quiet, wide market and disastrously slow around a breaking event. A maker should widen, reduce size or pull quotes when its data freshness and cancellation latency no longer match the market.

What changes when many agents use similar models?

AI agents can lower the cost of monitoring many markets and generate more continuous quotes. More competition can tighten spreads and deepen visible liquidity. That is the attractive case.

But similarity creates new fragility. Agents using the same news feed, model provider or risk trigger may update together. They may all cancel at the moment liquidity is most valuable. If they share the same mistaken fair value, they can reinforce a bad price until informed traders trade against them. If one agent reacts to another agent’s quote as though it were independent information, feedback loops can emerge.

This does not mean “AI manipulation” whenever an order book moves quickly. Similar behavior can come from ordinary arbitrage, shared public news or deterministic bots. Claims about agent coordination require evidence beyond a pattern on the chart.

The controls remain familiar:

• cap inventory by market, event and correlated exposure;
• enforce maximum spread deviation and order size;
• expire quotes before known catalysts;
• cancel everything when data, position or reconciliation checks fail;
• separate a forecasting model from deterministic risk controls; and
• stress-test simultaneous cancellations, partial fills and delayed updates.

Our defensive guide to prompt injection in trading agents adds another requirement: public market content must not be able to rewrite tool permissions or execution policy.

Can we tell which orders came from AI?

Not reliably.

An order book can expose prices, sizes and changes. A blockchain or API may expose addresses and transactions. Those observations do not prove whether a human clicked, conventional code executed a rule, an LLM suggested a trade or a company controlled several addresses. “Bot-like” speed is not a unique fingerprint for AI.

We should therefore analyse observable behavior—quote duration, depth, spread, inventory proxies where defensible and post-fill markouts—without assigning an identity or technology that the data cannot establish. Even a self-described agent should be treated as a claim unless its operation is independently auditable.

The Boxmining AI Trading Agent Scorecard follows the same evidence rule: we score declared systems under controlled tests; we do not label anonymous market participants as AI.

The market maker’s real scoreboard

A credible evaluation separates gross spread capture from inventory profit and loss, fees, incentives, adverse-selection markouts and operational failures. It reports fill ratio, quote uptime, average quoted spread, depth, inventory distribution, maximum drawdown, rejected orders and kill-switch events.

Polymarket’s liquidity-rewards documentation says current incentives favor qualifying resting liquidity, tighter quotes and two-sided depth, with market-specific settings. Those rewards can affect economics, but they can change and should be reported separately from trading performance.

If a strategy is profitable only after a temporary incentive—or only before realistic stale-fill losses—that distinction belongs in the headline result.

Read more practical market analysis in Technical Analysis.

AI market-making FAQ

What does a prediction-market maker do?

It posts buy and sell orders around an estimated fair probability, seeking spread or incentives while managing fills, inventory and changing information.

Can AI agents make markets?

Yes, agents can help monitor markets and estimate probabilities, but deterministic systems should enforce quote, inventory, permission and kill-switch limits.

What is market-maker inventory?

Inventory is the net position left after uneven fills. It creates directional risk and should influence subsequent quote prices and sizes.

What is adverse selection in prediction markets?

It occurs when a trader hits a stale or mispriced quote using fresher information, leaving the maker with a position whose fair value has moved against it.

Why does latency matter to a market maker?

A quote can become stale between observing an event and completing a cancellation. Fill and cancellation races can create unexpected positions.

What happens if many agents use the same model?

They may update or cancel together, share the same error and create feedback loops or sudden liquidity gaps around common news and risk triggers.

Can an order book reveal which orders came from AI?

Not reliably. Prices, sizes, speed and addresses do not prove whether a human, deterministic bot, language model or organization controlled an order.

How should market-maker performance be measured?

Separate spread capture, inventory profit and loss, fees, incentives, post-fill markouts, uptime, depth, drawdown, rejects and operational failures.

Do liquidity rewards guarantee market-making profit?

No. Incentives can change and may be smaller than adverse-selection, inventory, execution and operational losses. Report them separately from trading performance.

How should an AI market maker be tested?

Use replay and paper trading with realistic latency, queue position, partial fills, simultaneous cancellations, stale data and strict external inventory limits.

Sources and further reading

• Polymarket, Prices & Orderbook.
• Polymarket, Market Making.
• Polymarket, Real-Time Data via WebSocket.
• Avellaneda and Stoikov, “High-Frequency Trading in a Limit Order Book”.
• Glosten and Milgrom, “Bid, Ask, and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders”.

Risk disclosure: Market making and prediction-market trading can lose money rapidly through adverse selection, inventory imbalance, latency, contract-resolution risk and operational failure. Liquidity rewards and quoted spreads do not guarantee profit. This article is educational, not investment, legal or financial advice, and it does not identify any market participant as an AI agent. Use simulation and paper trading, verify platform rules and jurisdictional availability, and do not risk funds you cannot afford to lose.