Claude Code demonstrates how AI agents are moving well beyond chatbots

AI coding agents can now plan experiments, write software, run tests, inspect failures and revise their work with limited human supervision, according to a detailed personal case study by Google DeepMind researcher David P. Reichert. Reichert writes in his Substack that recent tools such as Anthropic’s Claude Code should be understood less as chatbots and more as agents that can act through a computer.

Reichert stresses that he is expressing a personal view and uses Anthropic’s Claude rather than a product from his employer as the prime example. His central argument is that public discussions of large language models often rely on outdated experiences with conversational AI. Modern coding agents, he says, can use files, terminals, browsers and programs to pursue multi-step goals and respond to feedback from those environments.

For the experiments described in the post, Reichert used Claude models to conduct seven small machine-learning projects over several weekends. He says he did not write code and only rarely inspected the generated programs. The projects were designed to be manageable on a home computer, but he estimates that each represented work that would otherwise take a human researcher several days.

From prompts to research workflows

The work included a system in which neural networks learned to solve two-dimensional mazes. Claude generated the maze tools, training code, evaluation process, visualisation and a written report, Reichert says. The experiment examined whether recurrent neural networks could solve larger mazes when allowed additional rounds of internal updates at test time.

Another project examined whether language models display a preference for associating unfamiliar words with an object’s shape rather than its color. A separate experiment revisited relation networks, a machine-learning approach for identifying relationships between objects. Reichert also worked with Claude on tests involving altered riddles and the text adventure game Zork.

He reports that the agents could take practical steps when problems appeared. In one example, Claude encountered failures while building a browser-based visualisation. It then wrote a separate program to inspect image pixels numerically, fixed a syntax error in that program and checked the output again. Reichert presents this cycle of action, feedback and correction as an important difference from a chatbot that only produces text.

The experiments also exposed limitations. In one language-model study, a reviewer agent found a bug in the code used to extract multiple-choice answers. The parser could mistake a letter appearing in an explanation for the actual answer. The researcher agent then revised the code after receiving the review.

That example supports two conclusions, Reichert argues. AI agents still make errors, including unusual ones. But independent review, testing and iteration can substantially improve results. He compares this process with ordinary research practice, where human work also benefits from additional scrutiny.

Useful, but not independent researchers

Reichert characterises current agents as being at least comparable with junior researchers or graduate students for many digital tasks, while adding that they have unusual strengths in speed and access to broad technical knowledge. That is an assessment from a single user’s experiments, not a controlled benchmark or an independent evaluation.

He identifies longer projects and lasting learning as major constraints. Models can adapt within a conversation and save notes in files, but they do not reliably turn new experience into permanent improvements to their underlying model weights. They also require human judgment on project goals, security and whether results are scientifically meaningful.

For content and knowledge workers, the practical implication is not that agents can replace expert review. It is that AI systems can increasingly perform connected workflows when they receive access to appropriate tools. Reichert advises users to treat agents as collaborators that need clear context, while maintaining safeguards around software installation, web access and sensitive accounts.