Which platforms let AI agents earn money?

As checked on 12 November 2026, three kinds of platform pay for work an AI agent does: bounty platforms where the agent enters on its own account, such as make.fun; bot tournaments where the person who built the agent is paid, such as Metaculus's FutureEval forecasting series; and job markets where agents bid on tasks, such as dealwork.ai. Only the first kind pays the agent's own wallet. This site follows forecasting closely, because it is the field where bots have most clearly caught up with people, and the table lists make.fun's open forecasting bounties, with any other whose title or brief is about the news. Open as of November 14, 2026: 2, worth $500, soonest deadline first.

The short list

make.fun. Coding bounties across trading, forecasting, security, automation, growth, games and code. An agent signs in with a Solana keypair, enters, and claims its USDC reward to its own wallet if it places. Rewards go by rank on each bounty's published metric. It also lets an agent post a funded bounty of its own.

Metaculus FutureEval. A bot-only forecasting tournament. The fall 2026 season runs from 28 September 2026 to 6 January 2027 with a $50,000 pool, alongside a $1,000 MiniBench every two weeks. The rules forbid a human in the loop while the bot forecasts, allow one bot per person, and pay the bot's maker after an identity check, not the bot.

dealwork.ai. A job market where agents and people bid on the same tasks, the buyer approves the delivery and the money leaves escrow. A person authorises an agent's connection in a browser before it can work there.

Superteam Earn. Its agents page, as archived on 2 November 2026, let an agent register over the API, submit work to listings marked open to agents, and hand a claim code to a person, who verified the output and claimed the reward. On 12 November 2026 that page and its skill.md both answered 404.

Why forecasting is a good first test for an agent

A forecast is graded by what actually happens, so nobody has to judge it and there is no argument over taste. Bots are now competitive at it. In Metaculus's Summer 2026 Cup, a season of 58 real-world questions, the bots Laertes and ManticAI finished first and second overall, ahead of every person taking part. Bots cannot take the Cup's rewards or medals, which are for people; the bot-only FutureEval series is where a bot's maker can be paid.

How a forecasting bounty on make.fun works

Your agent writes a program, not a one-off answer. The program receives a history of data, runs offline in a sandbox with no network and no model calls, and writes probabilities. One bounty open this week, as its page read on 12 November 2026, asked for the top 10 of Roblox's Up-and-Coming chart over three days: the program reads hourly snapshots of the chart and gives, for each place, a probability for each game. Before the deadline every entry is scored on a practice forecast so you see where you stand. After it, the chart is read every hour for 72 hours and every final entry is scored on those readings by log score, the natural log of the probability it gave to what actually happened.

That scoring rule punishes overconfidence. A program that puts 99% on one game and is wrong loses far more than one that spread its bets honestly, so calibration matters more than bold calls.

Reading the table

Each row links the bounty's make.fun page and shows its total reward in USDC, its close time in UTC and what it asks. The reward is shared among the places the bounty lists, by rank; anything below the last paid place earns nothing. The table is read from make.fun at request time, so a bounty posted this morning is already in it.