Twenty years of solvers

I have built optimisation software for twenty years. For most of that time, every improvement was handcrafted. You read a paper, found the place in the solver where it might fit, implemented it, ran the benchmark and stared at the numbers. A good idea took weeks. Most didn’t make it. Some that did were noise.

So I started keeping a ledger. Every idea got an entry: the hypothesis, what would make me stop, written down before the run, then the result and the verdict. Most entries say rejected. Those taught me more than the wins, because nobody forgets a success, but a failure gets tried again unless someone writes it down.

Then AI came in, bit by bit. First it completed lines. Then whole functions. Then I let agents run the whole lifecycle: take a ticket, research it, write the code, build, test, open the pull request. An agentic SDLC. Writing code stopped being the bottleneck. Knowing whether a change really helped became the bottleneck.

Then Andrej Karpathy published autoresearch. An agent edits a training script, runs a short experiment, keeps the change if the score improved, throws it away if not, and repeats. It inspired us because of what it really is: an optimisation loop. It searches over code the way a solver searches over plans.

That is the idea behind Argmax, and behind the name. In maths, argmax is the input that scores highest on a function. Here the input is a version of your code and the function is your metric. One level up, building a solver is itself an optimisation problem. We can’t try every version of your code, so we search. The moves are changes, proposed by a stochastic process: agents that research, design, write and test them. An agentic SDLC. And like any good metaheuristic, it lives or dies by its acceptance rule. Ours is a statistical test the agents can’t touch. Agents do the legwork. The ledger keeps them honest.

These are not generic coding agents. Everything twenty years taught me is built into them: how to read a solver, where the time usually goes, which kinds of ideas tend to fail, how to write an experiment down. The researcher agent thinks like an operations research academic and designs each change from methods the field already knows work. The implementer builds it the way a careful solver engineer would.

It runs on your infrastructure, with your model keys. The code, the pull requests and the research behind them are yours.

— Christophe Van Huele, founder