Replit launched Missions, a mode where you state an outcome, set a spending cap and choose which actions need your approval, and a group of AI agents works toward it in parallel. Replit's own test was to ask for a monitoring tool it could run itself. In two days, within a $3,000 budget, the agents produced dashboards, alerts and tracing, and Replit released a self-hostable version for anyone.
The contract is new, the swarm is not
Agents working side by side were already part of Replit's Agent 4 in March. What Missions adds is a contract written before any work starts: what counts as done, how much the agents may spend, and which steps wait for a person. A separate AI judge checks the result against those criteria, and if the money runs out first, the work stops unfinished.
That last rule is what makes the work priceable. Whoever signs knows the most it can cost, which is not true of paying contractors by the hour for a rebuild with no end date.
Why a monitoring tool was the right test
Replit could grade the new tool against the one it replaced. The original tool's own collector sent data into the rebuilt platform, so the agents had an answer key. To check it, Replit planted a database bug, and the new tool raised an alert in 38 seconds, against a goal of 60.
That points to where this kind of work gets cheap first: wherever a machine can check whether the job is done. Replacing an old reporting system, converting a pile of files into a new format, or rebuilding an internal tool whose outputs you can compare line by line all have that shape. Work where "done" is a matter of taste, such as a brand or a negotiation, does not.
What a buyer can do with this
Take a mid-sized distributor that pays a vendor $30,000 a year for software that turns its order data into weekly reports. The old reports are the answer key: a rebuilt tool has to reproduce the last three years of weekly figures to the cent. A mission capped at $3,000 with that test costs a tenth of one year's fee. The distributor may never run it, but asking the vendor at renewal what it would cost to leave is now a question with a real number behind it.
The limits are real. Replit's tests compared its tool with the original vendor's software in matched runs, not with the vendor's paid hosted service. The $3,000 buys the build, so it leaves out the staff time to set the goals and the work of fixing and securing the tool afterwards. And one company rebuilding its own product, with its own engineers writing the goals, is not the same as a shop owner doing it alone.
Where this leads
The subscriptions most exposed are the ones whose value is mostly code: monitoring, reporting, internal admin tools. Those with real value in the data they hold, a network of other users or a compliance certificate are much harder to copy. Vendors in the first group will answer with lower renewal prices and longer contracts, and buyers in that group will find the price list is now a starting point.
So at your next renewal, work out what test a replacement would have to pass to do the same job as the current product. If you can write that test, you have already worked out the price of leaving.