Coding agents improve quickly because code provides fast, objective feedback through compilation, tests, and runtime measurements.
2
Jay Hack argues that coding agents are generalist agents because code lets them create and use tools for many knowledge-work tasks.
3
The strongest agent products will combine access to context, many work surfaces, good economics, and a single system of record.
Summary
Jay Hack traces coding agents from early tools such as Kite and GitHub Copilot to Cursor and autonomous ticket-to-pull-request systems. He explains why code became a strong training ground: generated code can compile, pass tests, or be measured for performance, giving agents feedback they can act on. From there, he argues that coding agents are general agents because code is also the mechanism through which software tools interact with the world. The discussion moves to ClickUp, where Hack sees a chance to combine tasks, documents, chats, and other work data in one place. He is skeptical that fragmented tools such as Slack can give agents enough context. The conversation then covers AI-assisted science, model weaknesses in simple physical reasoning, world models, mechanistic interpretability, catastrophic forgetting, and agents that can run experiments or spend money. Hack is enthusiastic about the direction, but he is also direct about weak evaluation, unclear boundaries, and the difficulty of removing knowledge from capable models.
Coding products win when they arrive in the user's existing workflow
Jay Hack says early code-generation companies were often too early. Kite shut down even as GitHub Copilot gained attention because Copilot was integrated directly into VS Code and required little effort to try. Cursor extended the same pattern by adding a conversational sidebar, which became much more useful with Claude Sonnet 3.5. Hack expects other industries to adopt AI through similarly low-friction products. The important experience is that users can keep working while AI appears inside the tools they already use.
Verifiable feedback made code a strong training ground for agents
Hack says his early text-to-Figma and dashboard projects taught him that useful AI applications have verification built into the domain. A generated schema can be checked, compiled, and sent back to the model when it fails. Code has the same property. An agent can write code, run a unit test, inspect the failure, and try again. Performance work can also use measured runtime or CPU cycles to select better variants. Hack connects this loop to reinforcement learning from verifiable rewards, where the system can learn from outcomes that are easier to check than ordinary language responses.
Coding agents benefit from positive transfer into other knowledge work
Hack argues that coding is closely tied to reasoning, so improving a model at coding can improve its performance elsewhere. He also points out that tools used by agents are made of code. An agent that can write and execute code can create its own scripts and tools, which makes coding agents general-purpose systems in his view. He describes the shift from Codegen's ticket-to-pull-request workflow toward broader knowledge-work agents. The same basic loop can apply wherever an agent needs to act through software.
Agent products need context, work surfaces, and workable economics
Hack says future agent products will compete on access to context, the surfaces where people interact with agents, unit economics, and distribution. ClickUp appeals to him because it combines chats, tasks, documents, and whiteboards, while keeping organizational history in one place. He calls this idea "Cursor for your whole job." Generative interfaces could build a custom dashboard or JavaScript layer for a specific task. The tradeoff is that users must describe what they want and provide the right context, rather than simply consuming a fixed interface.
Fragmented data limits what general-purpose agents can do
Hack uses Slack as an example of the problems caused by fragmented systems. Codegen could not reliably answer why a decision had been made because a Slack integration could see only a limited slice of conversation history. He says general knowledge work has many more artifacts to reference than code, so first-party integrations and tagging are needed. He expects software platforms to converge because agents will increase the number of messages and tasks people receive. A single system of record could reduce the problem of deciding which inbox or application contains the relevant work.
Code is easier to evaluate than many tasks in the long tail
Hack says advertising, image generation, and other tasks are harder to evaluate because their success criteria are less direct than compilation or testing. An ad agent might use lower cost per click or cost per acquisition as feedback, but many other tasks lack a clear measure. He thinks evaluations are still useful for preventing regressions and checking long-context workflows, such as creating an image and uploading it to a service. He also expects basic model intelligence to improve quickly enough that some evaluations become obsolete within months.
AI systems are already being used around the frontier of science
Hack says language models are increasingly involved in mathematical and scientific work. He cites a Google DeepMind result involving a perfect score on the Putnam exam and discusses efforts to simulate cells. He also describes the possibility of using Claude Code to produce large numbers of empirical economics papers from public datasets. In biology and materials work, he mentions systems that connect models to laboratory experiments. Periodic Labs, for example, is described as giving Claude a tool for running experiments and measuring heat dissipation in semiconductor work.
Simple physical questions expose the limits of language-only models
The conversation uses the car-wash prompt to illustrate a model failure. Asked whether someone 50 meters from a car wash should drive or walk, GPT-5.2 reportedly recommends walking, missing that the car itself needs to be washed. Hack says this is not necessarily fixed by giving the model another software tool. He connects the failure to a missing world model. A system trained through robotics and interaction with physical objects could develop a stronger understanding of cups, cars, and other objects than a model that only predicts language.
Removing dangerous knowledge from a capable model is difficult
Hack says naive attempts to remove knowledge can make the model focus on the forbidden subject and strengthen its representation of it. He also raises catastrophic forgetting, where changing one part of a model may damage other capabilities. Even if a model could forget a biology concept, a sufficiently capable system might relearn it from other sources. Hack expects open-source models and information on the web to make permanent deletion especially difficult. He says the more realistic target may be removing the intention to misuse knowledge, though he says this has not yet been demonstrated.
"The solution for that is not necessarily code execution. I think actually if you just take a world model, if you have a robotics model that actually does interact with glasses and then you staple on a language model, this thing is going to understand the world better than you and I do."Jay Hack47:00
Who should watch
You are building coding agents and need to understand why compilation, tests, and runtime measurements make code easier to train and evaluate.
You lead an AI product team and are deciding whether to build a focused tool or a broader work platform with access to organizational context.
You want a grounded discussion of where current models fail, including physical reasoning, evaluation, interpretability, and forgetting.
