How to talk to an AI agent while it works

You have asked an agent to prepare a report. While it searches for information, you remember a detail: it should only use official sources. You type the clarification and press send.

That action leaves a question unanswered: will it read your message in time, or only once it has finished?

Hermes and OpenClaw offer modes for managing messages during a run.[1][6]

Codex and Claude Code also let you intervene while they work, but the shortcuts and the point at which the instruction is incorporated differ.[3][5]

Understanding that difference helps you correct course without restarting a task and save interruptions for when you need them.

Three intentions worth distinguishing

Steering, or steer. You want the work to continue, but with a correction: “Use official sources”, “Do not change the database”, “The recipient is a beginner”. The agent needs to receive it at a point in its execution where it can incorporate it. That point depends on the system.[1][5][6]

Queueing, or queue. You want to add work without cutting short the current task: “When you finish, prepare a summary”. Be careful with the name: a queue may deliver messages within the same turn or hold them for the next one. Knowing that a message is “queued” is not enough.[2][3][5]

Interrupting, or interrupt. The previous instruction is no longer valid, or you need to stop the run: “You are using the wrong document”. Each application offers its own cancellation controls.[4][5][6]

As a rule of thumb: steer to clarify; queue to add a later task; interrupt to abandon the current direction. If a sensitive action is involved, do not rely on a late clarification to replace permissions and prior approvals.

Hermes: choosing how to handle new messages

Hermes offers /busy in its interactive interface and gateway channels. The command shows or sets how messages arriving while it works are handled.[1]

• /busy queue: saves messages for a later turn.
• /busy steer: incorporates instructions during the work, after a tool call.
• /busy interrupt: redirects the active turn with the new message, preserving the work already done; to cancel it, use /stop.[1][2]

You can also express a specific intention without changing the entire policy:

• /steer Use only official sources.
• /queue When you finish, prepare a short version.

These commands let you steer a run or reserve work for the next turn.[1]

For example, choose queue if you tend to send several independent tasks. Use steer if you work conversationally and your messages are usually clarifications of the task in progress. Reserve interrupt for situations where you want to redirect the active turn sooner; use /stop to cancel it.

In messaging, check the bot’s acknowledgement. The Hermes documentation describes status messages indicating whether the input has been queued, incorporated as steering or has interrupted the work.[2]

Codex: Enter and Tab have different functions

In Codex CLI, the documentation distinguishes two ways to send text while the agent is working:

• Enter sends instructions to the current turn.
• Tab queues the message for the next turn.[3]

For example, if it is modifying an application and you remember that it must not add dependencies, type that restriction and press Enter. If you want it to review the documentation afterward, type that second task and press Tab.

The keymap in the official code assigns Esc to cancelling a running turn.[4]

These controls apply to the terminal interface. In the IDE extension, chatgpt.followUpQueueMode selects between queue and steer; the legacy value interrupt is interpreted as steer, not as cancellation. The default setting is queue. Cmd/Ctrl+Shift+Enter inverts the behavior for one message.[7] Do not automatically transfer terminal shortcuts to another surface.

Claude Code: queueing can steer the current turn

Claude Code lets you type while it is working. Pressing Enter adds the text to the message queue.[5]

What matters is when it leaves that queue: if tools are running, messages are delivered when they finish, before the next model call. They can therefore change the continuation of the same turn. If no tools are running, they are processed when the response ends.[5]

To redirect it, you can press Esc, type the correction and send it. The documentation notes that interrupting preserves the work and context accumulated up to that point.[5]

It also offers Ctrl+B to move running Bash commands and agents to the background, keeping the conversation available so you can continue working.[5]

Since version 2.1.275, Ctrl+Enter or Ctrl+X Ctrl+S lets you send queued messages immediately. Its effect depends on what Claude is running; it does not always amount to cancellation.[5]

The practical lesson is simple: do not read “queue” as a guarantee of a separate later turn. If you need a task to wait, say so explicitly: “When you finish the current review, prepare the summary”.

OpenClaw: steering, waiting or grouping messages

OpenClaw uses /queue to control message behavior during a run. The documentation consulted identifies steer as the default mode, including on Telegram.[6]

Its options include:

• /queue steer: steers the active run at the next safe interruption point.
• /queue followup: waits for the run to finish and then starts another turn.
• /queue collect: gathers several messages into a single later turn.
• /queue interrupt: aborts the active run and handles the most recent message.[6]

It also lets you adjust the wait for grouping inputs (debounce), queue capacity (cap) and overflow handling (drop).[6]

There is an important limitation: on Telegram and Discord, the current documentation says that collect behaves like followup, keeping turns separate. Do not assume it will group your messages on those channels.[6]

For someone who writes in bursts, collect can be convenient on supported surfaces. To keep tasks separate, choose followup. Before following an old tutorial, compare its modes with those documented for your version.

When this is useful

Imagine these situations, as working guidelines rather than exact equivalents between commands:

You are refining a report. The agent is already searching and you remember that it should exclude advertising. Steering lets you add the criterion without discarding the task.

You have a second task. While it prepares the report, you want to ask for a list of questions for a meeting. Saving that for later avoids mixing two different outputs.

You spot an error in the starting point. The agent is analysing the wrong case file. Interrupting makes more sense than letting it finish an answer you will not use.

You write in fragments. You send “prepare the summary”, then “in Spanish” and then “for someone with no technical knowledge”. A mode that groups messages may work better than treating them as three independent tasks, if your channel supports that grouping.

What these controls do not solve

An interruption should not be treated as an undo command. Claude Code, for example, preserves previous work when interrupted.[5] If files have been modified or information sent, check what happened before continuing.

Do not use these conversational controls as a substitute for cancelling a scheduled task, disabling its retries or stopping an external process, either. Those are issues that need to be checked and managed separately.

For sensitive work, establish the limits before starting: what the agent may modify, when it must ask for permission and which actions it must not take. “I’ll stop it if I see anything strange” is a poor strategy for an irreversible operation.

A simple way to start

During your first sessions, practise with harmless tasks. Ask for a draft, add a clarification while it is being prepared and check when the agent receives it. Then try reserving a second task. Finally, interrupt and review what work has been done.

You do not need to memorise every mode. You need to recognise your intention and know the corresponding control in your tool. Before pressing send, distinguish between “take this into account now”, “do this afterward” and “stop what you were doing”.

Scope note: this guide is based on official documentation consulted on 4 November 2026. It describes documented controls, not a functional test of every application or a check of their installed versions. The recommendations and examples are suggested uses proposed in the article.