Suppose you finish a task with Claude Code and a short quiz appears, asking what the agent just built and whether you understood it. You never asked for the quiz. Thariq Shihipar, who works on Anthropic's Claude Code team, uses this kind of example to explain Claude Mods. Mods is a system for customizing Claude Code, Anthropic's AI coding agent, from the inside. He walked through it with swyx and Vibhu, the hosts of Latent Space, in an episode published September 29, 2026.
According to Shihipar, Mods lets users customize the entire Claude Code harness. A harness is the software around the AI model. It gives the model tools, runs them, feeds the results back and shows the user what is happening. With Mods, users can change both how the harness runs and how its interface looks.
What is new compared with hooks
Claude Code already had hooks and plugins. Asked what Mods can do that those could not, Shihipar said the project was first called "function hooks" inside Anthropic. An ordinary hook links an event to an outside script: when X happens, run this script. A mod instead runs inside Claude Code's own TypeScript runtime, the program that is actually running the agent.
He said running in the same process gives a mod live information about the conversation, such as how many turns it has had, how many tokens it has used and what the messages say. (Tokens are the chunks of text that models read and bill by.) Mods can also attach to many more events. Because everything happens in the same process, a mod can start subagents: helper copies of the model that do a side job. It can read what those helpers return in a fixed, structured format. And a mod can change the interface, "which you can never do in hooks," he said. As an example of interface changes, he pointed to a Tetris demo shown inside Claude Code.
Anthropic's public proposal, dated September 3, 2026, describes the same design: TypeScript function hooks that can change how Claude Code runs and what it displays, in both the terminal and the desktop app. Mods can also be stacked, with each one wrapping the behavior of the next. The proposal says administrators can limit what extensions are allowed to do, and it gives auditing events and hiding sensitive content as example uses. A September 9 update to the proposal describes early access and plans to ship the system. In the interview, Shihipar said Mods works in the command-line and desktop versions and might come to Claude Tag, Anthropic's team-chat agent, in the future. Anthropic, he said, wants requests for more extension points.
How the automatic quiz works
Shihipar described the quiz mod step by step. At the end of each turn, the mod starts a "forked" agent: a copy of the current conversation that runs a small side request. The fork asks one question: has this task been completed? If the answer is true, the fork writes quiz questions and answers in JSON, a structured data format. The mod then reads the JSON and shows the questions just above the box where the user types.
Forking sounds expensive, but he called it "one of those unintuitive things." AI providers keep a prompt cache, meaning they store the already-processed part of a conversation so repeated requests can reuse it cheaply. A fork starts from exactly that cached conversation, so the extra question costs little. He still said the check is "slightly token intensive," because it runs after every turn. It is a lightweight classification, he said, that saves the user from remembering to ask: "You don't need to remember to do it."
This matches a pattern described in Anthropic's Agent SDK article, where subagents work in separate context windows and return only selected findings, not their whole working history.
Assumptions, next steps and a supervisor
Shihipar described other mods he is building on the same idea. One gives Claude a tool he calls "register assumption" (he said the name may change). Each time the model makes an assumption, it records it in a list, and the full list is shown when the work is done. This automates the habit of keeping implementation notes that show choices the model made on its own.
Another is a "next steps" mod. The conversation turned to what such a check should do: reread the original goal, ask whether the solution really met it, and suggest follow-ups as multiple-choice options. Shihipar said a next-steps mod could also recommend other skills. For example, it could suggest an explain skill after a complicated change, or point out that the user keeps asking for small fixes and might do better with a clearer prompt.
The benefit of building this as a mod, he said, is that it runs in a forked subagent, "so, it doesn't remain in the context afterwards." The main model keeps working while something like a supervisor checks the next steps from the side, and its notes do not fill up the model's working memory.
Routers, modes and the cost question
Shihipar is also building a model router, a mod that decides which Claude model handles each request. He said Anthropic does not route between models by default because routing is a hard problem and easy to get wrong: a router can send a hard problem to a model that is not up to it.
That raised a practical worry in the conversation: a mod that routes every query to a different model could quickly use up a user's subscription limit. Addy Osmani's cost guide, published September 25, 2026, explains why. Switching models starts over with an empty prompt cache, so a route that looks cheaper per token can cost more overall once the conversation has to be processed again.
The discussion also asked who Mods is for. Shihipar said power users, "but, like, the nature of Claude Code is that so many people are power users." Mods can be shared, so one person can build a good router "that doesn't break prompt cache all the time" and others can install it. He said Anthropic wants a skill for building mods that teaches Claude details like prompt caching, so that Claude can warn users who ask it to write one.
Mods can also build on one another. Shihipar described a mod that adds a mode selector at the top of the screen, where any other mod can register as a mode. The router could be one mode. An "artifact mode" could be another, in which Claude mainly answers through artifacts, Anthropic's persistent interactive documents. "The ability to create modes is in itself a mod," he said.
A preview of 'mutable software'
Shihipar called all this "a preview of" what he calls mutable software: applications that users can safely reshape to fit their needs, if the maker allows it. He said he hopes more apps do something similar, and suggested that startups ask Claude what an extension system for their own product would look like.
An objection came up in the conversation. Others have tried highly customizable software, and when users can do everything they tend to get confused. What usually works, the argument went, is one opinionated flow: a single well-chosen default way of working. The reply was that the opinions could themselves come as a skill that users install. Asked whether the design borrowed from JavaScript build tools such as Babel and Webpack, which also have plugins that stack, Shihipar said he is not deep in the technical details. He said the system was a collaboration between someone on the Bun team, which works on the JavaScript runtime, and someone on the Claude Code team.
If everything is customizable, what is Claude Code?
A host's question followed from earlier in the conversation, when "the bitter lesson" had come up in connection with harness engineering. The phrase comes from AI researcher Rich Sutton's argument that general methods using more computing power tend to beat hand-built cleverness. If much of this scaffolding disappears as models improve, and everything can be customized, what is Claude Code?
Shihipar said he was "misusing a little bit" the bitter lesson, which is about scaling and compute. He uses it as shorthand for the idea that "harnesses go out of date very quickly," and in ways that are hard to predict. Moving from chatbots to agents required entirely new tools. Now the model can change its own harness. Anthropic's own writing has examples of this churn. Shihipar's tool-design essay describes replacing a rigid to-do reminder system with a more flexible task system as models improved. A Managed Agents engineering post describes a context-management workaround needed for Sonnet 4.5 that was no longer necessary with Opus 4.5.
Shihipar argued that models are already far more capable than the average software task requires, and that the goal is still to deliver value to the user. He sees artifacts and mods as ways to spend that spare intelligence on keeping people informed and reaching the right result. But he drew a line. The core of Claude Code keeps getting more complicated. It needs a sandbox to operate safely, auto mode to handle permissions, computer use, MCP connections to data, and web search and fetch. "As the models can do more and more, the core harness has to be actually, like, quite complex and very secure," he said. "But then, like, how you interact with it can change quite a lot."
He expects models will eventually be able to write something like Claude Code in one go, and said they can already produce simpler harnesses in one attempt. He described a "barbell effect." For complex coding work, he said, use Anthropic's harness. For simpler or more specialized tasks, build your own lightweight harness on top of pieces such as Claude Managed Agents, which keeps the hard infrastructure in place while letting developers write a bare-bones loop for their task.