Ask on the Side

  • Desktop, mobile and web
  • All plans

What it does

Ask on the side opens a bubble where you ask a question about what an agent is doing, without the agent hearing it and without stopping it. The question runs off screen against a copy of the agent's conversation, read-only where the CLI offers a verified mode; the answer streams into the bubble and nothing enters the agent's history. It replaces the /btw some CLIs offer, without taking over the terminal. Use it to understand a decision ("why did you take that approach?"), check progress, or get a second opinion while the agent keeps working.

Video (1 min 6): Ask on the side, shown in the real app. A Claude Code agent keeps working while a question asked in the bubble gets its answer in about ten seconds.

Where to find it

  • The round Ask on the side button at the bottom right of the agent's pane, above the input box, next to the conversation view (Conversation View) and Read aloud buttons. The bubble opens upwards.
  • It is hidden for CLIs that cannot answer on the side.
  • The bubble header carries a video tutorial icon (a help icon with a small play badge) to the video, left of the close button.

How to use it

  1. Click the round Ask on the side button. The bubble's subtitle says where the answer comes from: "A copy of this agent, answering you" (the agent already has a session) or "A fresh session, reading this console" (no session yet, or a CLI that cannot copy it).
  2. Type the question in the field ("e.g. why did you take that approach?") and press the arrow Ask button. A waiting line shows what the CLI is doing ("Opening a copy of the session…", "Waiting for the model…") with a seconds counter, then the answer streams in, rendered in Markdown. The copy is made from the conversation the console is writing right now, under the account that console runs on. When that conversation is not on this computer (an agent running on an SSH host, a conversation started on another machine), a fresh session that reads the console answers instead, and a line above the answer says so: "The agent's conversation could not be copied, so a fresh session reading this console answered."
  3. Ask follow-up questions: the thread is kept, one thread per tab, named after its question. The + of the tab strip opens another side conversation (up to 5 per agent), which thinks in parallel with the others.
  4. The model chip in the bubble header opens the model picker (the same one as Change provider, without the provider part) and chooses which model answers. A new tab starts on a lighter, faster model (Sonnet on Claude, the Luna tier on Codex), unless the agent already runs on a model as light. If the conversation is too long for that model, the agent's own model answers and the bubble says so.
  5. On each answer: Copy, or Send to the agent, which only pre-fills the composer and never sends. When a question fails, Report this bug opens the Report a bug form already written up (provider, model, platform, the command that ran, what the CLI said): read it, add a word if you want, and send it. It becomes a ticket on the public AgentsRoom board, with its link, and you get an e-mail when the fix ships. Your question itself is not copied into the ticket, since its text is public.
  6. Drag the top edge or the top-left corner of the bubble to make it bigger; the size is remembered on your account.
  7. Closing the bubble does NOT cancel anything: the question keeps running, the round button spins while it thinks, then shows a badge with the number of answers to read until you open that tab. Closing a tab cancels its run.

Settings

None.

Agent tools (MCP)

None. The side question is a gesture of the user; agents cannot start one.

Providers

  • Runs on Claude, Codex, Grok, GitHub Copilot CLI, OpenCode, Mistral Vibe, Antigravity CLI and Kimi Code.
  • Read-only is enforced where the CLI provides a verified mode: plan mode on Claude and Grok, a read-only sandbox on Codex, the plan agent on Mistral Vibe. Other CLIs run without a special flag; since nobody is there to approve, a non-interactive run answers without touching the disk.
  • Model picker on Claude, Codex, Grok and Antigravity CLI; other CLIs answer with their default model.
  • Codex: the copy of the conversation needs codex-cli 0.148.0 or later (exec fork sub-command). AgentsRoom reads the installed version once per launch: on an older Codex the question runs as a fresh session that reads the console, and the subtitle says so. When the version cannot be read, nothing is degraded.

Mobile

Present, with the same two entry points as the desktop: the floating dock of the console (bottom right, above the input box) and the agent's actions sheet. The phone only sends the sentence; the desktop runs the very same side question, in the same tab, and streams the answer back rendered in Markdown. Tabs, model picker and badge are shared with the desktop: a question asked on the laptop shows on the phone and the other way round. Not on the phone: Send to the agent and Report this bug.

Limits

  • Runs on your own CLI account and spends your provider quota, unlike the Agent advisor which uses AgentsRoom's own allowance.
  • A thread is not persisted: once the console is closed it describes nothing any more.
  • One question runs per tab at a time; up to 5 tabs per agent.
  • Not offered on an agent started by a webhook trigger with restricted permissions (see Webhook Triggers).

Common questions

  • How do I check what my agent is doing without stopping it? Click the round Ask on the side button at the bottom right of its pane, type your question and press Ask: the agent keeps working and never sees the question. There is a short video: https://youtu.be/LuKI3uhX2wo
  • Does the agent notice? No. It carries on, undisturbed, and never sees the side question or its answer.
  • Can a side question modify my code? Not where the CLI offers a verified read-only mode (Claude, Grok, Codex, Mistral Vibe). Elsewhere the run has nobody to approve actions, so it answers without acting.
  • Which model answers? A faster, cheaper one by default (Sonnet on Claude). Change it in the bubble header; the agent's own model is marked in the list.
  • Does it cost tokens? Yes, like any turn. That is why it starts on a lighter model, and why closing a tab cancels its run. Closing the bubble does not: the answer waits under the badge.
  • Can I ask several unrelated questions at once? Yes: open a tab per question with the + of the bubble; each one runs on its own copy of the conversation.
  • My side question only said "failed". Fixed on 2026-10-18: a failing side question now shows the reason the CLI gave (for example "Not logged in"), so you know what to fix.
  • A follow-up failed with "No conversation found with session ID". Fixed on 2026-10-18 (Claude Code and Codex): AgentsRoom asks the question again once on a fresh copy of the agent, replaying the earlier questions and answers of the thread.
  • The very first side question failed with "No conversation found with session ID". Fixed on 2026-11-03: the copy was looked for under the account and the conversation the agent's settings named, while the console could be running on another account (a quota rule, an account changed since it started) or writing another conversation. The question now targets what the console really runs; Codex and Grok agents bound to an account get the same fix. If the conversation still cannot be copied from this computer, a fresh session reading the console answers, and the answer says so.
  • Why does an answer say "a fresh session reading this console answered"? The agent's conversation could not be copied from this computer (it runs on an SSH host, or it was started on another machine). The answer comes from a session that read the console, so it knows what was said there, not everything the agent has in mind.
  • On Codex the side question spun for five minutes, then expired without an answer. Fixed on 2026-10-30: a Codex agent in Manual permissions (it asks before acting) made the side question wait for an approval nobody could give. A side question now never waits for approval; a blocked command fails at once and the model answers.
  • On Codex it answered "unexpected argument" and exit code 2. Fixed on 2026-09-16: that was a Codex CLI older than 0.148.0. AgentsRoom now falls back to a fresh session; updating Codex restores the copy of the conversation.
  • Is there a video that shows it? Yes, about a minute, filmed in the real app: https://youtu.be/LuKI3uhX2wo