Working with your agent

hob is a workspace your agents can operate. An agent can arrange the context for a task, open a working preview, run a check, and point to the detail that needs your judgment. It can also use a web page while you watch and hand a step back to you. You spend less time moving information and opening tools on its behalf.

This connects the work from issue to review and commit. The agent brings the result and evidence to you in the same workspace. You direct the task, try the result, and decide what happens next.

The agent shows you something

Ask for the result where you can use it. A mockup belongs in a preview pane; a design note, PDF, or source file can sit beside the conversation. Commands and their results belong in a terminal you can inspect and reuse. An agent can also keep a previous file version when you ask.

Agents launched inside hob receive guidance for its CLI. For example, an agent can open the guestbook's HTML file in a live preview:

hob pane open web --file index.html

The preview reloads when the HTML file changes. An agent can add watches for related files with --watch, or open a URL for your project's development server. Browser previews explains file watching and server recipes; Files and previews covers the other kinds of context you can bring into a pane.

An agent can draw its explanation. A mermaid fence in a completed reply becomes a diagram. While the reply streams, you see its source. Ask for a diagram when a process is easier to see than to read:

The same viewer works in a Markdown file. See Diagrams in Markdown for its controls, source editing, and rendering limits.

The agent points at the detail

An attention link gives an explanation a precise destination. Instead of “look somewhere in the preview,” an agent can point at a web element, a file line, a terminal command, or a setting. hob brings the target into view and shows a temporary spotlight with the agent's note. Press Esc to dismiss it.

For a web element, the agent first reads the live pane with hob pane view. That returns node references and a snapshot fingerprint. The agent passes both to hob attention; hob checks the reference instead of guessing when the page has changed.

Attention appears as part of the agent's request, before the command returns. Its temporary card leaves a marker you can revisit. The attention link replays the pointer.

When the agent asks you a question

Some decisions need your judgment. The agent can present a question with suggested answers and a custom-answer field. Choose an option or explain what you want, then submit your answer to continue the conversation.

Questions, permissions, and plans serve different purposes:

• Questions ask what you want. Number keys select the displayed options; you can also write your own answer or skip a question.
• Permission requests ask whether an action may proceed. Review the action before choosing Allow once or Deny.
• Plan approval asks you to review a proposed approach. Depending on the backend and mode, you can approve, deny, or hand the plan off for implementation.

While a permission request is open, its card replaces the composer controls. Allow once answers this request; Allow always also enables Auto approval for the pane.

Hand off starts a fresh agent with the plan text. To continue the same conversation from a chosen turn, fork it.

The available controls depend on the agent backend and its selected permission or planning mode. Opening a preview or explaining a command does not replace the approval that an action needs. To change direction while a turn is running, use queueing or steering.

The agent walks you through it

Ask for a walkthrough when a change needs more than one stop: the behavior in the preview, the code that produces it, and the check that verifies it. The agent can offer a walkthrough link with ordered steps. You choose when to start it. Use Next to advance and Esc to leave.

You can mark up a surface for an agent, or ask the agent to leave notes for you. Mark up and annotate covers both directions, drafts, sent batches, and saved file revisions.

The agent arranges the work

An agent can name its pane for the current task, open supporting panes, and create a workspace for a separate workstream. Pane titles become a useful index as the project grows: “Add a counter,” “Review accessibility,” and “Check saving” tell you where to return.

You can also ask agents to work together. An agent can launch a delegate with an initial task, send a question to an existing agent pane, and link contributing conversations to an issue or commit. See Parallel agents for launching, messaging, and using @- to hand a pane over as context.

Messages between agent panes are delivered at turn boundaries. The receiving agent sends its answer back; the sender should finish its turn to receive it. Use clear, bounded assignments, especially when one agent is advising and another owns implementation.

An agent can also change a hob preference when you ask in plain language. The optional /settings prefix supplies command guidance and matching setting definitions.

Keep the decisions and evidence together

When an agent creates a commit, issue, or other supported artifact through hob, the result can link back to the turn that produced it. Open it later from History to recover the reasoning, or save useful context for another session.

The shared interface keeps you in the loop. A feedback draft waits for your review and Send action. An issue pane opens when you ask to view it. Your agent's managed guidance includes these collaboration boundaries alongside its CLI instructions. For reusable prompts and pane actions, explore Slash commands in the composer.

Next