Tasks and Plans

The Active Task tab and the History tab in the Build Agent panel work together but serve different purposes. Understanding the relationship between tasks and plans, and how the agent handles context in a task, helps Creators work with the Build Agent more effectively.

Tasks

A task is a conversation thread between a Creator and the Build Agent. Each task has its own history: every message sent, every response received, and every plan generated. Tasks are scoped to the module they were created in and persist across page reloads and sessions.

When a Creator opens the Build Agent panel in a module, the active task is where they left off. All previous messages and plans from that task are available in the chat history.

Tasks are isolated. Starting a new task clears the conversation. The agent begins with no knowledge of what was discussed in any previous task: no component names, no decisions made, no context from earlier sessions. Each task is a clean slate.

Task isolation surprises many Creators who expect the agent to remember what was built in a previous conversation. It does not. If a Creator needs the agent to understand what already exists in a module, that context must be provided at the conversation's outset.

Task and Chat Privacy

Tasks and chats are private to the Creator who made them. Other Creators in the same workspace cannot see another Creator's tasks, chat history, or plans. Each Creator in a workspace has their own isolated set of tasks and chats for every module.

Plans

A plan is the Build Agent's proposal for how to fulfill a request. When a Creator describes what they want to build, the agent generates a structured plan before doing anything. The plan lists exactly which Smart Components it will add, with their labels and Property IDs, and organizes complex requests into numbered steps.

A Creator must approve the plan before the agent builds. Selecting Build Plan executes it. To request changes, add inline comments on specific plan steps first, then select Request Changes. The agent incorporates the comments and presents a revised plan for approval.

Plans belong to tasks. Every plan the agent creates is associated with the task that produced it. One task can generate many plans, for example when a Creator iterates on a request or asks the agent to revise and resubmit.

How Tasks and Plans Relate

The relationship is hierarchical: a task is the conversation, and plans are outputs of that conversation.

Select the ellipsis button on any task row in the History tab to open the task actions menu. The menu includes a Download Debug option for support reporting and a Delete Task option.

Starting a New Task

Create a new task when the current conversation has run its course and the next piece of work is unrelated to it. A new task gives the agent a clean slate. Use one when starting a different feature, working on a separate section of the module, or resuming after a previous task grew too long.

Context lost when starting a new task:

• The entire conversation history from the previous task.
• Any context the agent built up through the conversation, including component names, design decisions, and requirements discussed.

Context retained: All plans created in previous tasks remain visible in the History tab.

When starting a new task that builds on previous work, open with a brief summary. Cover which components already exist, what they do, and what the new task should accomplish. The more specific the opening context, the more accurately the agent can proceed.

Providing Context Within a Task

Because no context carries between tasks, the quality of a task's opening message directly affects the quality of what the agent produces. Before describing what to build, consider including:

• A brief description of the module's purpose.
• The names and roles of components already on the canvas.
• Any naming conventions or validation rules that apply.
• The specific action the agent should take next.

Providing context at the beginning of a task is more reliable than relying on the agent to infer it from the canvas state alone.

How the Agent Manages Long Conversations

The build agent operates in a context window, a fixed limit on how much conversation history the agent can process in a single request. As a task grows long, the agent approaches the limit.

When the estimated token count for the next request exceeds 90% of the context window, the agent automatically compresses the conversation history. It summarizes older messages into a single compressed entry and retains the most recent conversation events in full. Compression happens transparently. Creators do not need to take any action.

Effects of compression:

• Long tasks might lose specific details from earlier in the conversation as compression occurs.
• If the agent references incorrect component names or repeats work already done, compression might have reduced the fidelity of earlier context.

If an error related to the context window displays, that error is not expected. Compression is designed to prevent it. Begin a new task and provide a brief summary of what was built so far to continue. Report the error through Unqork Support so the team can investigate.

Plan Statuses

Each plan in the History tab displays a status badge.

Editing a Plan

Plans cannot be edited directly. After the agent generates a plan, a Creator can:

• Approve it: Select Build Plan to approve the plan and begin execution.
• Request changes: Add inline comments on specific plan steps to mark what should change, then select Request Changes. The agent incorporates the comments and presents a revised plan for approval. If a Creator adds inline comments and then selects Build Plan instead, the comments are passed as context but the original plan structure is kept. Use Request Changes when the plan needs structural revision.
• Delete it: Select the delete button on a plan row. Deleting a plan closes its open tab and removes it from the current chat context.

Inline comments let Creators mark up individual steps with specific instructions before sending the plan back for revision.