Early versions of Oprel used a single AI model to handle everything: understanding the goal, planning the steps, and deciding each action. This worked for simple tasks but fell apart on anything complex. The model would lose context, repeat steps, or get stuck. So we rebuilt the execution engine from the ground up with two specialized brains.
The Strategist Brain
The Strategist is a high-capability model running with an extended thinking budget. When you type a task, the Strategist receives it along with a screenshot of your current screen and your task history. It does not execute anything. Its only job is to decompose the goal into phases: discrete chunks of work with clear success criteria.
Example: for "Prepare the Q3 report and email it to the team", the Strategist might produce:
The Strategist never touches the keyboard or mouse. It only thinks. Execution is handled entirely by the Executor.
The Executor Brain
The Executor receives one phase at a time from the Strategist, along with a live screenshot and the success criteria. It plans a batch of 15 to 25 concrete actions: mouse moves, clicks, keystrokes, scroll events, and executes them sequentially via the Ghost Hand.
After each batch, the Screen Understander analyzes the current state of the screen and the Verifier checks whether the phase's success criteria have been met. If yes, the next phase begins. If no, the Executor plans a correction batch.
The Screen Understander
Before each planning cycle, a lightweight vision model analyzes the current screenshot to produce a structured description of what is on screen: open windows, active fields, visible text, UI state. This gives the Executor precise context without sending raw screenshots to the planning model.
The Verifier
After each phase completes, the Verifier compares the screen state against the Strategist's success criteria. It answers one question: did this phase succeed? If not, it flags the failure mode and the Executor retries with a corrected plan. If the Executor gets stuck across multiple attempts, Stuck Chat activates and Oprel asks you directly in the chat panel what to do next.
Planning Mode
Before starting any complex task, Oprel enters Planning Mode. The Strategist asks clarifying questions one at a time, not a form, not a list, just a natural conversation, until it has enough context to build a reliable plan. This dramatically reduces mid-task failures.
The two-brain architecture means the system that plans is never distracted by execution details, and the system that executes is never overwhelmed by strategic complexity. Each brain does one thing extremely well.