Synkrasis Labs

The first half of AI was about method: architectures, objectives, training recipes, and the benchmarks built to measure them. That work largely succeeded. Language agents now reason over long horizons, call tools, and operate in environments that were out of reach a few years ago.

What it deferred is the harder half: defining what these systems are for, and establishing how we would know when they are competent. An evaluation that rewards a correct final answer says almost nothing about the trajectory taken to reach it, and a trajectory is what an agent is. An agent that arrives at the right answer is not the same as an agent that knew how to get there.

We treat evaluation as a first-class research problem rather than as infrastructure in service of someone else's model. Our work is full-path rather than final-state, interventional rather than observational, and grounded in real processes such as a working farm, a long-running conversation, or a constrained device at the edge, rather than in curated test sets. We build the systems we evaluate, on the principle that you cannot meaningfully instrument what you have never had to make work.

Synkrasis Labs is unaffiliated and self-directed. We hold no product roadmap and no institutional line, which leaves us free to pursue the questions we judge to be load-bearing rather than the ones that are conveniently fundable. Everything we produce is published.

Our reference point is Shunyu Yao's The Second Half, the argument that with methods now largely solved, the remaining work lies in problem definition and evaluation. We read it as a research agenda.