Agents in TIDE

The Agents section lists the project’s agents, children under their parents and a running plan’s agents under the plan, each saying what it is doing. Enter shows one live in the agent pane, where you talk to it as in its own terminal. The CLI side is in the Agents guide.

What you see

wdx tide · shop-api

Each row: its mark, its name, what it is doing and how full its context is. fix-auth asks a question, so its row reads yes/no. Click, scroll and type: it is live, and runs nothing.

The mark before the name: ● running, ? waiting for you, ○ exited, ✕ failed.
What it is doing, after the name: working, nothing when idle, or yes/no and decision when TIDE can read the question it asks (Claude Code for now), else waiting.
limited since 0.5.0, in red, in place of what it is doing when a usage limit its model draws on is reached and it is not working; its info tab says which and when it resets, and a folded parent counts it in N waiting. It goes by itself at the reset. See Agents for which limits are an agent’s.
Its context size at the end, 82k or 1.5M: green, then yellow from context.threshold.warning and red from context.threshold.critical.
The active row, the one shown on the right, starts with a green ▌; the cursor’s row is blue.
Children are indented under their parent with └. A parent folds with ← and opens with →; folded, it says how many below it wait for you.

An agent that asks a question is answered in its own tab, as in its terminal:

wdx tide · shop-api

Answer it without leaving TIDE; the row goes back to working. Click, scroll and type: it is live, and runs nothing.

Or answer every question of the project from one place, the prompts tab unreleased, without opening each agent.

i opens an agent’s info tab, what wdx agent get says:

wdx tide · shop-api

The info tab: Enter renames the agent. Click, scroll and type: it is live, and runs nothing.

C-q a opens the new agent form, C-q c the same form for a child of the agent under the cursor:

wdx tide · shop-api

Fill in what you want; empty fields take their defaults. Enter creates it. Click, scroll and type: it is live, and runs nothing.

Keys

Key	Where	What it does
`↑` `↓`	the list	move the cursor
`Enter`	an agent’s row	show it in the agent pane, the focus staying in the list
`o`	an agent’s row	show it and go to it
`←` `→`	a parent’s row	fold or open its children
`i`	an agent	its info tab
`a`	the project pane	the new agent form (`C-q a` anywhere)
`c`	an agent	the form for a new child of it
`r`	an agent, or its info tab’s `Enter`	rename it
`g`	an agent that ended	start it again, in the same session
`R`	an agent	a new session that reads its history in brief, after a confirmation
`x`	an agent	remove it, after a confirmation
`-`	an agent’s row, or its tab	close its tab

In the agent’s own tab, every key goes to the agent; use C-q and a letter there.

How it behaves

The form asks in one box what wdx agent create takes as flags: Name (empty for a generated one), ACA (← → among those installed and enabled, shown with two or more), Model (the ACA’s models; empty for your own setting), Worktree (empty for where it would start; a new name creates the worktree), Permissions (full or ask), and Prompt (one line, its first task). Tab moves between fields, Enter creates it, Esc cancels.
A new agent becomes the active one, live on the right.
Talking to it is typing in its tab: what you type goes to the AI CLI, and Shift+Enter and other keys reach it as they would in its own terminal. Scroll back with the wheel or PageUp.
Questions turn the row amber, its mark ?, until answered; a folded parent shows N waiting. With the agent’s question read, the info tab and wdx agent get show it with its options.
Removing or restarting asks first, saying what runs; R on a busy agent says what it is doing, and y goes on.
Agents that ended keep their row, their last screen and history, read-only, until removed or started again with g.

See Worktrees and diffs for an agent’s worktree, and Docs and plans for a plan’s agents.