--- title: "Cerevisor 4.0: the human step, where an AI workflow stops and waits for a person" slug: cerevisor-v40-human-step date: 2026-10-28 excerpt: "Cerevisor 4.0 adds the human step: a point in an AI workflow where the run stops, shows a person what the agents prepared, and carries on down the path of the button that person clicks. It waits as long as it takes, even in scheduled runs." featured_image: "https://bbtxujdxvidaghmhxkqs.supabase.co/storage/v1/object/public/generated-images/blog-1793776974846-cerevisor-v40-human-step.webp" featured_image_alt: A person reviews a document at a desk between incoming and outgoing stacks of work. author: Kinan Hamwi author_url: https://www.linkedin.com/in/kinanhamwi/ canonical_url: https://cerevisor.com/blog/cerevisor-v40-human-step updated_at: 2026-11-04T07:22:55.257038+00:00 --- # Cerevisor 4.0: the human step, where an AI workflow stops and waits for a person By the Cerevisor team. Checked on 28 October 2026 against the source code of Cerevisor 4.0.0. TLDR Cerevisor 4.0 adds the human step: a card in a workflow where the run stops and waits for a person. A panel called Your input shows what the AI agents prepared, the person answers and clicks a button, and that button decides which step runs next. It waits as long as it takes, also in scheduled and Hands-free runs, and costs nothing while it waits. A proposal for a client is ready at 7:40 in the morning. A research agent gathered the numbers overnight, a writing agent turned them into three pages, and a pricing agent prepared two options. Nothing has been sent. The run is sitting on one amber card that reads Waiting for you, and the laptop shows a notification: Your input is needed. One click on the card, a look at the two options, a click on Send option B, and the run carries on to the agent that formats and files the final version. That pause is the headline of Cerevisor 4.0. Cerevisor is a desktop app for people who do demanding work on a computer with AI. An agent here is an AI program that does one job, such as researching, writing or checking. A workflow is a set of agents that work toward one result, laid out on a canvas where each agent is a card. Until now, a workflow ran from start to finish on its own. With 4.0, a workflow can stop at a chosen point and ask a person. Every change is listed in the [Cerevisor 4.0.0 release notes](https://cerevisor.com/releases/v4-0-0). This post spends most of its time on the human step, because it changes what kind of work can be handed to agents at all. --- ## What a human step is in an AI workflow A human step is a card where the run stops for a person. It sits in the workflow like any agent card, but it is not an agent. It calls no AI model, uses no tools and has no files of its own. The agents before it do the work and hand it over, the person decides, and the agents after it carry on with that decision. Each human step has a name and a question, up to 20 fields to fill in (short text, long text, a number, a choice from a list, yes or no, or a date) and between one and six buttons. Each button is a possible answer. A new step starts with two: Continue and Stop here. Key Insight The human step is not a confirmation box bolted onto an agent. It is a place in the workflow that belongs to a person, and the button that person clicks steers what happens next. --- ## How the human step works, step by step - **Add the step** Open Manual controls at the top of the canvas and choose Add Human step, or ask the AI Builder (Cerevisor's built-in assistant that builds workflows from a plain request), for example: "After the research step, add a step where I pick which leads to contact, with the buttons Send and Skip." - **Connect each button to what should follow** Draw connections from the human step to the next steps, then open a connection and choose under Follow this button which button makes it run. A connection set to Whatever is clicked runs on any answer. A step on a path that was not chosen is skipped, and its card reads Path not taken. - **Run the workflow** The agents before the step do their work. When the run reaches the human step, it stops, and a panel called Your input opens with the question, what the agents handed over, the fields and the buttons. - **Answer, or come back later** Fill in the fields and click a button. Close for now closes the panel while the run keeps waiting, and the footer reads Your input is needed until the step is answered. - **The run carries on** The button picks the next step, and the answer is what the next agents read. Afterwards the card shows which button was chosen. Required fields must be filled before the step sends, and Cerevisor checks every answer before accepting it: the button must be one of the step’s own buttons, and each field must match its kind. --- ## A human step that waits as long as it takes, even in scheduled runs The part that matters most for real work is the waiting. There is no time limit. As the [release notes](/blog/cerevisor-v38-local-ai-agents-lasting-conversation) put it: > "It waits as long as it takes, in scheduled and Hands-free runs too. A waiting step calls no model and takes no place among the agents that may run at once." Cerevisor 4.0.0 release notes, October 2026 Hands-free is the approval mode in which agents work without asking along the way. A human step still stops there, and it never guesses an answer. Because a waiting step asks for no model and holds no place among running agents, a run parked overnight on a decision does not block other workflows from running. A desktop notification says Your input is needed, and clicking it brings Cerevisor to the front with the panel open. A paired phone with the Cerevisor companion app gets the same notification. --- ## The page: showing the agents’ work the way a person reads it By default a human step shows a plain form, which is always enough to answer it. A step can also have a page: a small screen the AI Builder designs for that workflow, such as a table of leads to tick or two drafts side by side. It is saved in the workflow’s folder and opens inside Cerevisor, never in a web browser. Two safeguards keep the page honest. A page cannot answer by itself: an answer only counts right after a person’s own click or key press inside it. And if anything other than the AI Builder changed the page file, Cerevisor shows the plain form instead, until the person reviews the page with sample data and clicks Use this page. When several human steps wait at the same moment, they appear one after another. Settings, Admin, Show human steps side by side puts them next to each other. > A page cannot answer by itself. The decision always takes a person's own click. --- ## Where to use human approval in AI agent workflows The human step is built for the moments in knowledge work where a person has to be the one who decides. Five that fit it well: - **Approving a draft before it goes out.** A writing agent prepares a client email, a newsletter or a report. The human step shows it, with the buttons Send and Revise. Send leads to the agent that formats and files it; Revise leads back to a rewriting agent with the notes typed into a long-text field. - **Choosing between options an agent prepared.** A planning agent drafts three versions of a project plan or three pricing options. The page shows them side by side, and the button clicked decides which one the next agents build on. - **Adding information only a person has.** A number from last week’s call, a client’s preferred tone, a date that is not written down anywhere. A field on the step collects it, and the agents after it read it as part of the answer. - **Signing off a budget or a client deliverable.** Before a cost estimate or a finished deliverable moves on, the run stops for an explicit Approve or Reject, and the card records which button was clicked. - **A check inside a scheduled or hands-free run.** A Monday morning report workflow can run on a schedule, gather and draft everything, and then wait at one step for a person to look before anything is sent. If a repeating schedule comes round while the earlier run is still waiting, Cerevisor skips that start rather than stacking up runs. --- ## What has not been tried yet with human steps Two limits from the release notes are worth knowing. A page made by the AI Builder on a real model has not been tried yet, so the plain form is the proven path for now. And the phone notification arrives only with the updated companion app, and only while that app is open on the phone. --- ## Also in Cerevisor 4.0 - **Memory in layers.** Memory is now one line per fact, in three layers: About you, My business, and a page for each area of work. Nothing is learned without a click: what Cerevisor picks up from runs and chats waits as a suggestion with Accept, Edit and accept, and Reject. - **Runs that check before they start.** A run checks what its skills need before it starts. A run that leaves loose ends now ends as Finished, needs attention, in amber, with a list of what the agents could not check. - **Webhooks.** A workflow can start from a web address on the same computer, and can tell another app what happened when it is done. A failed send is tried again after 1, 5 and 30 minutes. - **Interrupted work comes back.** After a crash, one Interrupted work window lists every run and chat reply that was cut off, each with its own Resume. - **GitHub Copilot as a provider.** An agent can now run on GitHub Copilot through the Copilot command-line program, under its own Preview tile. --- ## Try the human step in Cerevisor 4.0 The quickest way to see it is to take a workflow that already writes something, add a human step before the last agent, keep the buttons Continue and Stop here, and run it. The [Cerevisor download page](https://cerevisor.com/download) has 4.0 for Windows, macOS and Linux, and existing Windows and Linux installations get it through the built-in updater. The full list of changes, including every known limit, is in the [4.0.0 release notes](https://cerevisor.com/releases/v4-0-0). #### Sources - [Cerevisor 4.0.0 release notes](https://cerevisor.com/releases/v4-0-0) - Cerevisor, 2026-10-28