descartes

the problem

ai got really good at answering and never learned to ask. when a model doesn't actually know something it doesn't stop, it just gives you the most confident sounding guess. that's the root of hallucination. and it's way worse inside agents, because a wrong assumption there doesn't just give you a bad sentence, it ships bad code or takes a wrong action you have to undo.

what it does

descartes is a doubt layer. before an agent acts on your prompt it doubts every decision, then doubts its own doubts, then answers them. but only from real evidence. your actual code, cited sources pulled live, or a question back to you. anything it can't ground it flags instead of inventing. it loops until a full pass turns up nothing new, then hands you a doubt hardened plan and a short list of the calls only you can make. the point isn't more questions, it's the fewest sharp ones, with zero unexamined assumptions sneaking into the work.

why it's the best of both parts

the mcp half is the reach. it drops into claude code as a server so the doubting happens right where your agent already works, before it edits or acts, with your real code as ground truth.

the asker half is the brain. a real interrogation engine, sharp question moves, grounded answers, a convergence stop, and a panel of models where the disagreement between them is the signal for what's actually uncertain. it knows what to doubt and when to quit.

one without the other is a demo. together it's a layer your agent actually thinks through.

what we used (and how)

claude / claude code. the whole thing runs as an mcp server inside claude code. claude is the reasoning that does the doubting, and the pre edit hook is what makes it doubt before the agent acts, not after.

fireworks. the multi model panel runs here. we send the same doubt across several different model families through one fireworks key, and where they disagree is exactly where the real uncertainty is. that disagreement is our detector. fireworks is also where the fine tuned asker is hosted.

exa. the grounding. when a doubt needs an outside fact, exa deep search pulls real sources with citations and confidence, so descartes answers from something real instead of guessing. low confidence gets marked unknown, never asserted.

hud. where we proved it. we ran a real eval on hud showing an agent that asks beats one that guesses on held out facts, asked 0.80 vs blind 0.40. that's the number behind the whole thesis, measured on hud, not claimed.

openrouter. fallback routing and the gemini image generation for the pixel art descartes assets.

tracks

agentic collaboration. this is the core. descartes is a layer agents collaborate through, it makes one agent doubt and ground its decisions before acting, so it stops acting on wrong assumptions. that's collaboration between the agent, its own doubts, and the human on the few calls only they can make.

see it

on a frontier topic where a normal model bluffs, descartes shows you what's actually supported with sources, what's unknown, and what needs you. that refusal to assert what it can't ground is the whole idea. running live, and droppable into your own agent today.