Anthropic Turns Claude Code Into a Platform: Claude Mods Flip the Agent Harness

Anthropic is turning the Claude Code harness into a programmable layer. Mods (Claude Mods) are TypeScript or JavaScript plugins that run inside Claude Code, in the CLI and in Desktop. They subscribe to events (a prompt, a tool call, interface rendering), and they can observe those events, change how they are displayed, or replace the default behavior entirely. In Latent Space, Dan McAteer calls this an early form of the Attention Interface: a layer that decides when the agent interrupts you, what it shows, and which decisions it expects specifically from you. For several weeks the author tailored Claude Code to his own engineering workflow, and he reports concrete numbers on both the benefits and the costs.

Six mods already inside Claude Code

Anthropic itself built part of the product out of mods. You can see them via /plugin under Installed → Built-in:

• agents-md — loads AGENTS.md as project instructions;
• diff — takes over the /diff command and draws its panel;
• plugin-authoring — gives Claude a skill it uses to write mods from a plain-language description;
• sec-default — protects organization settings from mods the user installs;
• telemetry — sends analytics from Claude Code and the built-in mods;
• you-should-know — a side agent that shows a note above the input line when it notices something you may have missed. It is the only one of the six that is off by default.

The You Should Know case is telling. The author asked the agent to simplify noisy UI text in a personal project. The main agent trimmed the text on the Compare screen, built the project, and took before-and-after screenshots. The side agent noticed that CompareView.swift and roughly 45 other files contained uncommitted changes. Reverting the demo changes would have wiped out the old unfinished work as well. Its advice: commit or stash first.

The author wrote the second mod himself. Next Steps Supervisor runs after a turn in which the agent edited files, ran commands, or made 8 or more tool calls. It delivers a verdict with a list of concerns, such as shortcuts taken, and suggests a next step as a button that sends an instruction to Claude Code. The mod is part of ModSmith, the author's open plugin with a set of mods and a template for creating new ones.

Three ways to allocate attention

McAteer compares Claude Mods with two products that tackle the same problem:

• ChatGPT Intelligent UI. On October 7, OpenAI released GPT-6 in ChatGPT together with this interface. The model generates charts, maps, checklists, forms, and mini-tools such as a savings calculator right in the response. The model picks the interface from OpenAI's component library, and the only setting named at launch is reducing visuals. Settings are minimal, and the target user is not an engineer.
• Cursor PR Routing & Approval. It assigns reviewers based on code ownership and commit history, and it can automatically approve low-risk PRs. The decision rests on a risk score against a set threshold, Bugbot findings, Security reviewers, and policy files in the repository. If a PR changes the policy, the agent reads it from the base branch, so a PR cannot rewrite the rules to approve itself. Cursor's documentation states plainly that the feature "does not replace a full review." The team writes the rules, but the interaction flow is fixed.
• Claude Mods. The human and the agent build the interface itself in code and can hook into any point of the loop, from prompt submission to the end of a turn.

The author's formula: ChatGPT and Cursor give you an Attention Interface that someone else designed, while Claude Mods let you design your own. But the responsibility for making it useful falls on the user.

Noise, tokens, and trust

The author lists three downsides outright:

• Noise. Since installation on October 15, 99 of 111 turns met the trigger conditions of Next Steps Supervisor. When a mod speaks almost every time, you start scrolling past its verdicts, and the point is lost. In one session it marked a task "Solved" and immediately listed two gaps and three shortcuts.
• Cost. Each check forks the conversation and rereads it. Most of it comes from the prompt cache, but the mod's README estimates a check at 5–6K input-token equivalents on a 50K-token session, and it costs more when the cache has gone stale.
• Trust. A mod is code running with your permissions: it reads and writes files, sees every prompt, and spends your usage limit. Anthropic's documentation warns that mods are not isolated in a sandbox, and the built-in ones keep running even when you launch with --safe-mode.

One extra upside: you can fix your own mod. If the supervisor is too noisy, /supervisor off disables it, and you can rewrite the trigger conditions.

Where to start

McAteer proposes a simple test for any mod: did it show you something you would have missed, make a decision easier, or let you act without unnecessary output? If not, it is decoration. The practical order is this: notice friction in your work with Claude Code, describe the mod you need in plain language, and let the built-in plugin-authoring skill write it. Third-party mods are worth reading before you install them, and their trigger conditions are worth narrowing so tokens go to real risks. The author's idea of the next step: the harness itself will read session transcripts, find the places where you keep correcting the agent, and propose a mod for them.