Steer Your AI Agent Mid-Run: Message Queue, Chat Forking and Goal Mode in Jobbit
Steer an AI agent while it works: Jobbit's message queue, chat forking from any message, and the Auto, Plan and Goal execution modes explained, with a worked example.
Chatbots wait for you. AI agents do not: once they have a goal they plan, use tools and keep going, sometimes for an hour. That autonomy is the whole point, and it creates a new problem. What do you do when, three minutes into a build, you realise you gave the wrong brand colour? Or when a research run is heading somewhere useful but you want to explore a second direction without losing the first? Jobbit's latest release answers with three controls: a message queue that lets you talk to a working agent, chat forking that branches any conversation from any point, and execution modes (Auto, Plan and Goal) that set how independently the agent should work.
This guide explains each control, the situations it is for, and how they combine with Jobbit's other features such as the agent browser and Projects. It is written for people who use agents for real work rather than demos; the basics of how agents plan and act are in what is agentic AI?.
The message queue: talk to an agent while it works
In most agent products the composer locks while a run is in progress. You wait, the agent finishes, and only then can you say "actually, use the other logo". Jobbit keeps the composer open. Type while the agent is working and your message joins a queue that the server holds, so it survives a page refresh or a phone going to sleep. Messages go out in order as the agent reaches a point where it can take them, and the queue shows which one is going out next.
What it is good for:
- Corrections. "Navy, not black." "Use the 2026 price list." The agent picks the correction up at its next step instead of at the end.
- Additions. "Also add a contact page." "Include Scotland in the comparison." Extra scope without restarting.
- Sequencing. Queue three follow-ups while the first task runs and go and do something else; they execute one after another.
- Answers in advance. If you know the agent will ask which of two options you prefer, tell it before it asks.
Up to ten messages can wait at a time, and a queued message can be removed before it goes out. Because the queue lives on the server, the agent decides the exact moment to apply each message, which keeps corrections from landing mid-action. When the agent is waiting for a decision from you, the chat says so, and your next message answers it directly.
The queue pairs naturally with the live browser view: watch the agent in Jobbit's agent browser, see it heading the wrong way, and type the correction without stopping the run.
Chat forking: branch a conversation from any point
Sometimes the right move is not to correct but to split. You have a good research thread and want to try a different angle; you have a build you like and want to test a riskier design without touching it; a colleague wants to take a copy further. Forking creates a new chat that starts from the original, from a message you choose or from the whole conversation, with a "continued from" line pointing back to the branch point so either chat can jump to where they diverged.
- Fork from a message. Pick the reply that did the work you want to build on and branch there. Everything after it stays in the original chat.
- Fork the whole chat. From the chat header or the sidebar, copy the full conversation and continue.
- Follow the thread back. Both chats keep the link, so you can always see where a branch came from.
- Inside projects. Forks stay inside the project, with the same files, instructions and secrets, so a colleague who joins a project can fork your chat and carry on with the full context. Projects are explained in Jobbit Projects.
Forking is how you run experiments with an agent the way developers run branches with code: cheap to create, easy to compare, no risk to the original.
Execution modes: Auto, Plan and Goal
Next to the model button in the composer sits the execution mode, which tells the agent how much initiative to take:
| Mode | What the agent does | Use it when |
|---|---|---|
| Auto | Chooses the best approach for your task | Most of the time; the agent decides whether to plan or act |
| Plan | Creates a plan before taking action and shows it to you | Big or ambiguous jobs where a wrong start would be expensive |
| Goal | Keeps working until your request is done | Well-defined outcomes you want finished without check-ins |
Plan mode is the safest way to start anything large: an app with several workflows, a research project spanning many sources, a launch with copy, images and automations. The agent lays out the steps and asks its questions before spending credits, and you correct the plan while corrections are free. Our guide to writing prompts for AI agents recommends asking for the plan first; this mode makes that the default.
Goal mode is for the opposite situation: the outcome is clear and you want the agent to pursue it to completion. The chat shows the agent working towards your goal and reports when it is not there yet, and the agent keeps going through failures and retries until the request is done or it genuinely needs you. It suits builds with a clear definition of done, migrations, data clean-ups and long research jobs, and it works best on the stronger model tiers described in choose your AI model in Jobbit.
Auto stays the sensible default. The agent reads the request and decides whether a quick answer, a plan or a long run is appropriate.
Working with a long run: context, notifications and files
Three smaller features make long runs comfortable:
- Context meter and compaction. The composer shows how much of the model's context the conversation is using. When a chat gets long, compact the context: the agent summarises what matters and continues with room to think. If the agent is waiting for a decision, resolve it first and then compact.
- Browser notifications. Turn them on and your browser tells you when a run finishes or the agent needs you, so you can leave the tab.
- Workspace files. Everything the agent creates in a chat appears in the workspace file explorer, where you can search, upload by drag and drop, rename, duplicate, move and download a selection as one archive, on desktop or phone.
Together with voice input in the composer (record, and Jobbit transcribes), these turn a chat with an agent into something closer to managing a colleague than operating a tool.
Start a build in Plan mode on Jobbit, correct the plan, switch to Goal mode, and queue your refinements while it runs. Start free.
A worked example
You ask Jobbit to build a booking site for a yoga studio.
- Plan mode. The agent proposes pages, a schedule model, a booking flow and the email confirmations, and asks whether classes have capacity limits. You answer and remove a page you do not need.
- Goal mode. You switch modes and let it build. The workspace panel shows files appearing; the agent browser opens as the agent clicks through the booking flow to test it.
- Queue. Halfway through you remember the studio uses a specific green. You type the hex code; it goes out at the next step, and the agent restyles without restarting.
- Fork. The site is live. You fork the chat to try a members-only area with a different pricing model, leaving the original chat clean for the version the studio approved.
- Notification. You close the tab; your browser tells you when the fork is done.
Every step used the same chat, files and project. The steps that used to need a developer's attention, restarting, branching, testing, happened as sentences.
Why these controls matter
Autonomy without steering is a gamble; steering without autonomy is a chatbot. The message queue, forking and execution modes let you hand an agent a job the size of a project while keeping the two things a manager needs: the ability to correct early and the ability to explore safely. As agents take on more, described in the future of AI agents, these become the core skills of working with them.
Frequently asked questions
Can I send a message while the AI agent is still working?
Yes. In Jobbit the composer stays open during a run. Your message joins a server-side queue of up to ten messages and goes out in order when the agent reaches a point where it can take it. You can remove a queued message before it goes out.
What does forking a chat do?
Forking creates a new chat that starts from the original, from a message you choose or from the whole conversation. Both chats keep a link to the branch point, and inside a project the fork keeps the same files, instructions and secrets. Use it to try a second direction without touching the original.
What is the difference between Auto, Plan and Goal mode?
Auto lets the agent choose the best approach. Plan makes it write and show a plan before acting, which suits large or ambiguous jobs. Goal makes it keep working until the request is done, which suits clear outcomes you want finished without check-ins.
What does compacting the context mean?
Long conversations fill the model's context window. Compacting asks the agent to summarise what matters so far and continue with more room, which keeps long runs coherent. If the agent is waiting for your decision, answer it first, then compact.
Do these features work on the free plan?
Yes. The message queue, forking, execution modes, context compaction, notifications, voice input and the workspace explorer are available on every plan within its credit and concurrency limits. Free accounts run two tasks at a time; paid plans run more, as described on the pricing page.
Give an agent a real job and keep your hands on the wheel. Start free on Jobbit.