Connect over one API key and doloop runs a loop around your AI. It produces, a fixed check holds the result to your standard, and nothing passes until it meets it. You write the standard once. The loop runs on its own after that.
Ask one AI to grade another and the grade bends: to flattery, to a confident wrong answer, to a crafted instruction buried in the work. The accept-or-reject in this loop comes from a fixed check you control. Same input, same verdict, every run. There is nothing to persuade.
Five signed messages between the side that holds the standard and the side that does the work.
The job to do and the standard it has to meet, signed by your side so it can't be swapped in transit.
While it runs, doloop sends a steady signal so your side knows to keep waiting, and can tell waiting apart from a stall.
The output comes back with a record of what changed and a check that it rebuilds the same way twice.
Your standard decides. A rejection points at the exact line that failed and why, not a vague score to argue with.
Approve how a case should resolve, and the loop applies that decision itself the next time it appears. It gets cheaper the longer it runs.
Every message in the loop is signed by the side that sent it. Tamper with one and the signature breaks. Two machines can trade work across a network and each can trust what the other said, with no person in the middle vouching for it.
So long as it takes an API key from you, it runs in the loop.
Hold a change to the rules your codebase keeps, and merge on a verdict you can replay.
See Code →Every value pinned to its spot on the page, or the cell stays empty. Nothing invented.
See Documents →Score a draft on the axes of your voice and name the one change that moves it toward yours.
See Writing →Your agent, your model, your key. doloop is the loop around it, not a model that stands in for it. Connect over one API key. Your agent does the work; your standard runs as a fixed check on our side and decides when it is done. The loop hands back an accept, or a rejection pinned to the line that failed, so your agent can fix it and go again with no one in the middle.
The loop is live. Keys go to design partners today: tell us the standard your agent has to meet, and we hand you a key and the loop.
Talk to us →The loop clears each pass on its own. You write the standard, and from then on you look at the exceptions, the handful of results that could not meet it. The rest go through without you.
A model cannot reliably grade its own work. The whole reason the loop holds is that the standard is a fixed check, not another opinion.
Read the thesis →Routes you to a real doloop page, asks when your question is ambiguous, or tells you when there is no answer. No model runs on the answer path, so it cannot invent one.