EXPERIMENT 04 · HOW ACTIONIST WORKS
How Actionist actually works
A perception–reasoning–action loop, a shared map of how tasks get done, and agents that delegate.
The loop
Actionist does not script a task ahead of time — it observes the screen, decides the next move, takes it, and checks what changed.Perceive
Actionist captures a screenshot of the current screen. That image is the ground truth — what the user would actually see, not an assumed page state.
Reason
A large action model reads the screen and the goal, then decides the single next step. It plans against what is visible, not against a brittle pre-recorded macro.
Act
Actionist executes the step through real mouse and keyboard control — clicking, typing, scrolling — the same inputs a person would use.
The Action Tree
Every completed task is a route on a shared map. The next agent doesn’t relearn it — it reads it.The Action Tree grows every time a task succeeds: each run contributes the concrete steps that worked, and overlapping paths reinforce the most reliable route. Because less than a fraction of the web is exposed through public APIs, an API-only agent simply cannot reach most software. Actionist works on the interface itself, so the map covers the real surface of an app — not just the slice a vendor chose to document — and that coverage compounds with every user.Multi-agent delegation
A single run is rarely a single job. Actionist splits work across specialists and brings the pieces back together.The orchestrator decides which sub-agent owns each part of the task, hands off context, and lets each one call tools through MCP. Results merge back into one coherent outcome the orchestrator returns.One real run
Here is the same architecture as a single end-to-end task, traced step by step.For your role
- Developer
- Enterprise buyer
- Community
The loop is extensible at every seam. Add capabilities as skills, connect external systems over MCP, and compose sub-agents to keep responsibilities small and testable. Because Actionist acts on the interface, you are not blocked waiting for an API that does not exist.
Ready?
See the loop run on your own screen.
Open the dashboard, describe a task, and watch perceive–reason–act–verify in motion.