Trajectory (Training)

A trajectory is the complete, gapless record of a path to a solution. It captures which task everything began with, which steps followed afterward, and what came out at the end. You can think of it like a logbook: not just the destination is noted down, but every single leg of the journey there. For a computer program that handles a task on its own, such a log records, for example: which file it opened, which command it tried, which error message came back. It is precisely these logs that are now being collected on a large scale in order to let computer systems learn from them. The technical term originally comes from physics, where a trajectory describes the flight path of a body.

Why final results alone reveal too little

Earlier language models learned almost exclusively from texts that contained the finished answer. A textbook shows the correct solution, but rarely all the dead ends before it. A model that only ever sees such final results learns to phrase things well, but not to work well. It knows what a correct solution looks like, but not how to arrive at it.

For tasks with many intermediate steps, this is exactly what matters. A programming assignment consists of searching, testing, finding errors, and correcting them. Anyone who only knows the finished code cannot imitate this process. Trajectories close this gap because they supply the intermediate steps as well.

On top of this there is an economic point. Freely available text from the internet has by now largely been exhausted. Good trajectories, by contrast, are scarce and must be laboriously generated or recorded by experts. That is why they regularly turn up in business news as an expensive and sought-after commodity.

How a solution path becomes training material

Trajectories are created in two ways. Either you record humans at work, for instance developers fixing a bug. Or you let the model itself work through tasks thousands of times over and log all the attempts. The second route is cheaper and delivers huge quantities, but the quality varies greatly.

The records are then evaluated. For a programming task, an automated test checks whether the code runs in the end. Trajectories with a successful outcome count as positive examples, the others as negative ones. The model is then adjusted so that it more frequently chooses sequences of steps drawn from successful paths. This procedure is called reinforcement learning: the system is rewarded for good outcomes instead of copying a prescribed answer.

A well-known problem here is credit assignment. If a path with thirty steps fails, it is unclear which step was to blame. Sometimes only the third step was wrong and the rest was clean. Research is therefore concerned with evaluating individual steps instead of just the final result.

Trajectories in agents and product news

Trajectories are most visible in so-called agents. These are AI systems that do not merely respond but independently operate tools: search engines, terminals, calendars. Coding assistants such as Claude Code or GitHub Copilot work this way. Their capabilities stem to a large extent from millions of recorded workflows.

The principle also shows up outside of software. Self-driving cars learn from driving records, robots from movement sequences that humans have demonstrated to them. In all these cases, the path is the learning material, not the destination.

Anyone reading reports about AI companies will therefore often come across data providers who pay experts to record such workflows. A typical mistake is to equate trajectories with ordinary training data. The difference lies in the structure: trajectories have a fixed temporal order and an end that can be evaluated.