Tasks & Deep Scan
Key takeaway: Deep Scan is a read-only AI audit that opens your project's real working directory, reads the code, config, dependencies and logs, and returns a short list of concrete tasks. Each task carries one of seven categories, an urgency, and a ready-to-paste instruction. The list is kept between five and eight active tasks, and tasks are never deleted.
Most task lists are made of things you had to think of yourself. In Heygents the list writes itself: a Deep Scan opens your project's real folder, reads the actual code, config, dependencies and logs, and comes back with a short list of concrete things worth fixing or building next.
What the Tasks tab is
Every project workspace has a Tasks tab in its main tab strip, alongside Info, Plan and Automations. It holds that project's AI-generated backlog: a grid of cards, each one a specific piece of work with a category, an urgency, an explanation of why it matters, and a ready-to-paste instruction for actually doing it.
Two things make this list different from a normal to-do list. First, nothing on it was invented in the abstract - every task comes from something the scan genuinely found in your project. Second, the list is deliberately short. Each scan keeps the active backlog topped up to at least five tasks and caps it at eight, so you are always looking at a page you can act on rather than an archive you scroll past.
Deep Scan: how the list is generated
The Deep Scan button starts a read-only audit inside the project's real working directory. It is a full AI run, not a static analyser: it reads source files, checks dependencies, looks at logs, and probes the live site or database where that makes sense. It works across four broad fronts - security and secrets, reliability and operations, performance plus SEO and growth, and code quality and data model.
Progress streams live while it works, and the run is detached from your browser, so you can navigate to another project, close the tab or switch to your phone and the scan keeps going. When it finishes, the new tasks appear on the Tasks tab.
Every task is scoped to the project itself - its own code, config, dependencies, routes, schema, assets and deployment. Shared-server and general infrastructure concerns are deliberately excluded, so you do not get a backlog full of "patch the OS" and "harden SSH" items that have nothing to do with the app you are working on.
How Deep Scan grounds ideas in your Goal
On the Plan tab each project has a free-text Goal: what this project is actually for and where it is heading. This is not decoration. Whenever a Goal is set, the scan receives it as context and grounds every Development and Upgrade/Feature idea in it.
The practical difference is large. Without a Goal, feature suggestions are generic best practice. With a Goal, they are proposals that move your product toward the thing you said you were building - so the ideas read like something a collaborator who understands the project would suggest, not a checklist. If your feature tasks ever feel off-target, the fix is almost always to write a sharper Goal and re-scan.
Re-scanning and scoped scans
On a project that already has tasks, running the scan again is a Re-scan: it re-checks each existing task against fresh findings and marks the ones that are genuinely fixed as completed, then tops the list back up. You do not have to tell Heygents what you finished - it notices.
The Re-scan button opens a small composer so you can aim the run. Tick one or more task areas to look at only those, pick which AI provider runs it, add free-text instructions, override the model, or customise the audit brief. A scoped run only creates tasks in the areas you ticked and leaves earlier tasks outside that scope untouched. Leave every area unticked and you get the full all-areas deep scan. It is the same form used to configure a scheduled Scan automation, so what you learn here transfers directly.
Task categories
Each task carries exactly one category from a fixed set, shown as a colour-coded chip on the card:
- Security - secrets, exposure, auth and permission problems.
- Performance - speed, payload size, queries, resource use.
- Development - net-new functionality that builds toward the project's goals.
- Upgrade/Feature - improving or extending something that already exists.
- SEO/Marketing - discoverability, metadata, growth surfaces.
- UI/UX Design - interface, layout, clarity and usability.
- Maintenance - stability, cleanup, dependency and technical-debt work.
Because the set is fixed, every task always matches a filter, even when the AI phrases a category loosely. Clicking a card's category chip filters the grid to that category, and clicking the active one again clears back to All.
Filters and task counts
Above the list, filters sit in two distinct groups. On the left, category tabs: All plus one pill per category actually present, each with its colour, icon and a live count. On the right, an independent status filter: All, Active, Completed, and Ignore/Postpone when anything is parked there, again with live counts. The two groups combine freely, and both reset when you switch projects.
Task counts travel outside the Tasks tab too. The Home dashboard ranks projects by open-task count in its My Tasks card and shows each project's open count in the projects list, and on phones the Tasks tab carries a green badge for tasks solved in the last 24 hours. Open means every task that is neither completed nor ignored.
Solving, approving and dismissing tasks
A task moves through open to in progress to completed. There is an optional intermediate state, awaiting approval, used when an automation has finished a task but its rule requires human sign-off - the card then shows Approve and Reject buttons, where Approve completes it and Reject sends it back to open.
Cards distinguish Auto-completed (an automation finished it unattended) from a human Completed, so you can always tell what was done for you and what you did yourself.
Separately, Ignore/Postpone is the manual dismissal. Pick it from a task's status control and the card dims and sinks to the very bottom, below the completed ones. It drops out of the category counts and the Active filter, and automations skip it entirely because they only ever pick up open tasks. It is not a completion, it carries no credit, and you can bring it back at any time through the same control.
One rule underpins all of this: tasks are never deleted. There is no remove button. A re-scan that finds a task fixed marks it completed rather than dropping it, and every project keeps at least five tasks on the list. Your backlog is a record, not a disposable list.
The explainer on task cards
Every task card carries a small What does this mean? button. Clicking it opens a popover with a short, plain-language explanation of that specific task: what the problem actually is and why it matters, in ordinary language rather than technical shorthand.
The explanation is generated on the first click and then cached on the task itself, so every later click is instant. The provider and model behind it are configurable per touchpoint in AI settings, under the Task explanation entry. The explainer is scoped to task cards only; it does not appear elsewhere in the interface.
Task workspaces and AI runs
Each task includes a ready-to-paste technical instruction for carrying out the fix, so a task is a starting point for work rather than just a note. From a card you can hand the task straight to the AI chat, where it opens as its own conversation named after the task - a small workspace dedicated to that one piece of work, running in the project's real directory. Because the chat is named after the task, it is never auto-renamed to a generated subject, and it stays easy to find in chat history later.
When an automation fixes a task, its full transcript is kept: the AI's notes and every tool it ran. The card then shows an AI run button that opens a full-screen replay of that run, formatted like a chat. The marker survives a later human approval, so a completed task keeps a permanent, readable record of exactly how it was done.
Let the backlog write itself
Set a clear Goal, run a Deep Scan, and let Heygents tell you what your project actually needs next.
Open Heygents →Frequently asked questions
How often should I run a Deep Scan?
Manually whenever you have finished a chunk of work and want the list re-checked. If you would rather not think about it, create a scheduled Scan automation and let Heygents refresh the backlog on its own. A repeat scan re-checks each existing task against fresh findings and marks the genuinely fixed ones as completed before topping the list back up.
Can I delete a task I disagree with?
No, and that is intentional: there is no remove button and tasks are never deleted. Use Ignore/Postpone instead, which dims the card, sinks it below the completed ones, drops it from the category counts and the Active filter, and hides it from automations - while keeping the record in case you change your mind.
Why do my feature ideas feel generic?
Almost always a missing or vague Goal on the Plan tab. Development and Upgrade/Feature tasks are grounded in the Goal, so a specific Goal produces specific proposals that move your product toward the thing you said you were building. If your feature tasks ever feel off-target, write a sharper Goal and run the scan again.
Does a scan interrupt my work?
No. Scans are read-only audits and they run detached from the browser, on your own machine under the companion agent, so you can keep chatting, browsing files or switching projects while one is in progress. Progress streams live while it works, and because the run is detached from your browser you can close the tab or switch to your phone without stopping it.