Matt

On Matt's first day of building the agent, the model cost only a bit over eight cents, and the server costs a fixed few dollars per month.

Matt pointed out that the horizontal-axis labels on one of the benchmark comparison charts were generic names rather than specific benchmark names, and that Kimi K3.1 ranked first on every item, questioning whether the chart was self-made.

Matt is the developer of Docs Agent, an open-source documentation tool built on LangGraph that can automatically analyze GitHub repositories, find documentation gaps, and generate fix PRs.

Matt believes that if you are trying to put a product into production or use it in an enterprise, evaluation is very important; but for independent developers, evaluation is worth the investment only when the agent really needs to perform well in a production environment.

Matt mentioned that if the model gets stronger tomorrow or the testing framework changes, the evaluation system also needs to iterate accordingly.

Matt suggests that people without a technical background learn the brakes rather than the accelerator first: at the start they should watch the agent's every move and understand each of its actions and the capabilities of its toolchain, then gradually expand to multi-agent setups, custom skills and external data source integration.

Matt puts forward a view contrary to the mainstream narrative: do not give the model the grandest tasks, but rather the most boring and trivial work (such as drafting replies and handling miscellaneous chores), because such tasks make output quality easy to measure and errors easy to correct.

Matt (founder of aihero.dev) has roughly ten years of hobbyist plus eight years of professional coding experience, long believed that large models writing code was just a bubble, and after a certain model capability leap point came to believe he would never write a line of code himself again.

Matt believes there is a severely underestimated learning curve with agents, and that stronger models will not eliminate this curve but only hide it deeper.

Matt says he now runs four agents simultaneously handling long-term tasks across different projects, with daily commits pushed to GitHub reaching 200 to 300, whereas in his hand-coding days 20 a day was considered good.

Matt believes the codebase is the agent's working environment, and the environment determines the output: putting a powerful model into a garbage codebase only yields garbage results.

Matt designed a skill called Grill Me that has the agent instead relentlessly interview the user, forcing out requirements, boundaries and implicit preferences.

Matt observed that the top-performing students are without exception excellent communicators who write clear and detailed descriptions when submitting code changes; those with poor output often are not lacking in technical skill but are unable to explain the problem clearly to the agent.

The flight price tracking task that Matt demonstrated took 30 to 45 minutes, with the agent automatically connecting to the internet, building the tool workflow, creating a skill and setting up a scheduled task.

During the livestream, when Matt manually sent a guest email invitation, he misspelled Alex's address, and a few minutes later a personal assistant proactively popped up to point out the spelling error and suggest following up with a correction.

In the demo evaluation, Docs Agent's documentation quality score was low and exposed a rendering prompt error, giving Matt a data basis to judge whether the code needed improvement or the evaluation method needed adjustment.

Matt once tried to organize a custom video editor project into a massive specification document and then have the agent generate the code, which lasted several months with unsatisfactory results, and he ultimately abandoned the specification and turned his attention back to the code itself.