On this page
  1. What a card contains
  2. Today: permission approval
  3. With proactive agents
  4. Three shifts
  5. Design principles

An attention card is a request by an agent to borrow a person’s attention once.

What a card contains

  • Context: which agent, on which machine, for which task.
  • Evidence: command output, exit codes, diffs, monitoring data. It comes from the kernel’s structured record, not from the agent’s own account.
  • Options: approve, reject, modify, later, save as a rule.
  • Consequences: which files change, which commands run, which networks are reached, and whether it can be undone.
  • Expiry: when the card lapses, and what happens then.

A card can be trusted because of the kernel:

  • Evidence comes from the structured record.
  • The policy engine turns one decision into a rule.
  • Snapshots provide undo.
  • State sync lets a person handle the card on any device.

An interface alone cannot provide these. Without the kernel, a card has only what the agent says about itself.

Today: permission approval

An agent is blocked and waits for a person to approve an action, such as running a command, changing a file or reaching the network. This is passive. If the agent does not ask, the person does not know.

With proactive agents

Agents start to work on their own initiative. They watch CI, logs, machine state, dependencies and inboxes, and prepare a plan when they find a problem. Cards grow from “the agent is blocked” to “what the agent found, what it wants to do, and what it has done.”

The examples come from the field notes.

Card type Raised by Content Example
Proposal An agent that found a problem I noticed X and prepared plan Y, verified in a sandbox or a dry run. Approve and it runs. “spark-a and outside nodes keep using the relay, because port 41641 collides with office-a. Change it to 41642. This needs sudo, and tailscaled drops for a few seconds.”
Finding or alert An agent that found a problem Nothing to do yet, but you should know. With evidence, and why it concerns you. “Free memory on spark-b is down to 2 GiB. memguard kills processes below 1 GiB.”
Grant A person or an agent Instead of approving one by one, a standing permission with a scope, a quota and a term. “Allow automatic restarts of DeepSeek on spark-a and spark-b, at most 3 a day. Stop and notify me after 3 failures.”
Clarification or preference An agent that needs what only a person knows A choice between two or a few options. The answer goes into memory. “KV cache at 1.2G or 1.8G? The first is steadier, the second keeps more prefix cache.”
Handoff An agent that cannot do it It needs a person’s hands: a sudo password, 2FA, a cable, a phone call, a physical act. “Please run this sudo command: edit fstab so that swap persists.”
Recap An agent, on a schedule What it did on its own within its grants. Each item can be undone. “In the past 8 hours I handled 12 things: 1 restart, 2 disk cleanups, and more.”
Conflict Agents in conflict Two agents want the same resource or edit the same file. “DeepSeek and the voice service both want the Spark’s GPU memory. Running both will run out of memory. Which one?”
Environment change An agent that detected a change The network, the place or the device changed, and some abilities are affected. “You joined in-flight Wi-Fi: about 940 ms round trip, UDP 443 blocked, Google times out in Chrome. Turn QUIC off for now?”
Opportunity An agent that saw a pattern Low priority. Can be collected and read later. “After every deploy you check 3 endpoints by hand. Make it an automatic check?”

Three shifts

  1. From approving one by one to grants and recaps.
    • The person’s role moves from operator and approver to supervisor and governor.
    • The weight of a card moves from “approve this once” to “set the rule” and “check the result.”
    • Trust is calibrated step by step through recaps. An action that is always approved can be proposed for a grant.
  2. From agents waiting for people to scheduling people’s attention.
    • Every card has an interruption cost, a value, an urgency and a reversibility.
    • The system decides whether to interrupt now, hold the card for a digest, finish within a grant and report afterwards, or drop it.
    • It also picks the channel: a terminal pane, a desktop notification, a push to a phone.
  3. Cards carry preconditions, checked again at approval.
    • A proactive card is based on the state at one moment. The person may handle it hours later.
    • At approval the preconditions are checked again, as in optimistic concurrency control. Does the port still collide? Is the machine still online?
    • If the state has changed, the card is invalidated or refreshed.

Design principles

  • Evidence before conclusions. A card holds output and data that can be checked, not only the agent’s summary.
  • Impact stated plainly. What changes, whether it can be undone, and who is affected.
  • One key is enough. The most common choice is the most visible, and batch handling is easy.
  • Decisions accumulate. Any decision can be saved as a rule with one key. Rules can be viewed, changed and revoked.
  • Do not interrupt by default. What can wait goes into a digest. What fits within a grant is reported afterwards.
  • Any device will do. On a phone, a person can at least understand the evidence and decide.