Evidence at a glance
The Harness Turns Toward Its User
Dan McAteer of Latent Space describes Claude Mods as plugins that run inside Claude Code, aimed at developers who want to tailor its CLI and desktop interface to their own engineering workflows. He uses them to change how Claude Code behaves and presents its work, and calls this an early form of an “Attention Interface.”
That framing identifies a change worth testing. A traditional agent harness primarily connects a model to a machine: it accepts prompts, invokes tools, and edits files. Claude Mods lets the user also shape how the agent presents its process, when it interrupts, and which decisions it hands back to a person. McAteer predicts that harnesses may shift from machine-side execution toward platforms for interaction between people and agents. That remains his thesis, not an established industry trend.
Event Hooks Change the Handoff, Not the Model
A mod consists of TypeScript or JavaScript functions that respond to events such as prompts, tool calls, or interface renders. It can observe an event, change how it is presented, or take over the default handling. Mods are therefore closer to middleware embedded in the agent runtime than to another panel that simply displays results.
For an engineering workflow, the key is that signals in an agent’s trace can be turned into actions for a person. Instead of leaving tool-call records in a log, a system could surface a warning near the prompt when a condition is met, or offer a clickable next step. This does not automatically make the agent better at execution. It changes when a person receives information, how they intervene, and whether they can continue the task directly from that intervention.
Useful Alerts Point to a Concrete Risk
McAteer’s example of You Should Know shows why this kind of handoff might matter. He asked Claude Code to simplify copy on the Compare screen of a personal project. The agent edited a file, installed a new build, and took a screenshot, making the task appear complete. A side agent then warned him that the edited file and about 45 other files already had uncommitted changes. Discarding the work from the demonstration could also have erased earlier unfinished work.
This is not an independent evaluation, but it makes the intended value concrete: surface state that is easy to miss in the main agent’s execution trace, so a person can act before work is lost. McAteer also built a Next Steps Supervisor. After file edits, command execution, or a threshold number of tool calls in a turn, it checks whether the goal is complete or something may have been missed, then offers suggestions in the interface. One mod says, in effect, “do not overlook this risk”; the other tries to turn a check into a usable handoff.
Alert Frequency Is a Product Trade-Off
The Supervisor figures are better treated as design clues than as proof of effectiveness. McAteer says that, among 111 Claude Code conversation turns he recorded after installing it on October 15, 99 met its triggering conditions, including a threshold of at least eight tool calls in a turn. Such frequent triggering suggests that a check can cover many work turns. It does not show that every alert is accurate or useful, or that the mod actually prevents errors.
Cost also shapes the experience. McAteer cites an estimate in the ModSmith README of roughly 5,000 to 6,000 input-equivalent tokens for one check in a session of about 50,000 tokens. Costs can be higher when the cache has expired. If alerts are frequent but their value is inconsistent, users may start skimming or ignoring them. An Attention Interface should be judged not just by whether it can produce alerts, but by whether its thresholds, false-positive rate, and per-check overhead fit the task.
Programmability Expands the Governance Surface
Making the human-facing layer programmable also gives plugins access to real work state. The research material notes that Mods do not run in a sandbox and can read and write files, inspect prompts, and consume usage. Third-party mods therefore need code review before installation. The same capability can warn about uncommitted changes or take over default handling. Convenience and risk come from the same mechanism, not from separate features.
There is also an unexplained discrepancy in the built-in plugin lists. The article names six: agents-md, diff, plugin-authoring, sec-default, telemetry, and you-should-know. The Anthropic repository README lists four built-in source directories. The available material does not say whether the difference reflects a version change, the scope of published source, or something else, so neither list should be treated as a definitive account of all built-ins. For teams, a practical starting point is to identify recurring human checks with clear consequences, define when an alert should fire, assess whether it is actionable, and review plugin permissions and maintenance. Customization is not a substitute for governance. It brings governance into the runtime.