The Antigravity AI agent — how it plans and executes

The agent sits at the centre of the Antigravity surface. Understanding how it works — not just what it can do, but the specific sequence it follows and the points at which you remain in control — is the most useful thing you can learn in your first week with Antigravity. This page covers the Antigravity plan-execute-review loop in full, explains what triggers a checkpoint, describes the interjection interface, and documents the handoff mechanism for multi-agent pipelines.

The plan step

When you submit a task description, the Antigravity agent does not start browsing immediately. It enters a planning mode and constructs a structured task plan. The plan is a numbered list of steps, each of which specifies the action to take (navigate, click, read, write, run), the target (a URL, a CSS selector, a file path, a shell command), and the expected output that would allow the Antigravity agent to consider the step successful.

The plan appears in the Antigravity plan review panel before anything executes. You can read each step, edit the action or target for any step, add a constraint (for example, "skip this step if the error rate is below 1%"), reorder steps, or delete steps entirely. You can also add new steps that the agent did not include. If the plan looks wrong in structure rather than in detail, you can reject it and describe the task again with more context; the Antigravity agent will produce a revised plan.

The review phase is not a rubber stamp — teams that use Antigravity in production have reported that ten minutes of plan review saves two hours of cleanup from an agent that made an incorrect assumption in step three. The plan is the cheapest place to catch a misunderstanding.

The execution loop

Once you approve the plan, execution begins. The Antigravity agent works through the plan steps sequentially, though it may generate sub-steps dynamically if a page turns out to be more complex than the plan expected. You watch a live output stream in the run panel showing the agent's current reasoning: what it is reading, what it decided, what it is about to do next.

The execution loop handles common web patterns without additional configuration. Paginated results are followed automatically until the agent has the data it needs or reaches a declared limit. Form validation errors are read and the Antigravity agent retries with corrected values up to three times before surfacing a checkpoint. Redirects and OAuth flows are followed within the declared credential scope. File downloads are saved to the run's local workspace and logged in the artifact.

The agent does not attempt to solve captchas. When it encounters one, Antigravity surfaces a checkpoint immediately and waits for you to solve it manually. After you solve it, execution resumes from the same step. This is an explicit design choice: captcha solving would require either a third-party service or a bypass, both of which conflict with the Antigravity security model.

Checkpoints and interjection

A checkpoint is a deliberate pause in execution triggered either by the Antigravity agent (when it encounters a decision it cannot make confidently) or by you (by pressing the interject button at any point during a run). When a checkpoint fires, the run panel shows the agent's current state: the last step it completed, the decision it is stuck on, and a specific question for you to answer.

Interjection via the interject button gives you a text field to type a correction, a clarification, or a new instruction mid-run. The Antigravity agent reads your input, adjusts its context, and continues from the current step with the updated understanding. You do not need to stop and restart the run; interjection is always mid-flight. The correction and the agent's response to it are recorded in the Antigravity artifact timeline as a paired event, so reviewers can see exactly what human input occurred and when.

Watch mode is the layout that makes checkpoints comfortable. In watch mode the run panel is full-height on the right side of the screen, showing the agent's output stream and the checkpoint interface. The browser pane on the left shows exactly what the agent is seeing and doing. Most users run in watch mode for the first few runs with a new task type, then switch to background mode once they are confident in the plan.

Review and the artifact

When a run completes — successfully or after a stop — the artifact becomes available. The artifact timeline shows every step the agent executed, with timestamps, the agent's reasoning for each decision, and the inputs and outputs. You can expand any step to see the full HTTP response, the file diff, the test output, or the screenshot. The timeline is scrubable: drag the playhead to any point and the browser pane shows the page state at that moment.

Annotation tools let you mark individual steps as approved, flagged for review, or incorrect. Flagged and incorrect steps generate structured feedback that you can attach to the next run of the same task type, helping the agent avoid the same mistake. Teams have adopted the annotation workflow as their standard review process for AI-assisted changes, treating the artifact timeline the way they treat a pull request diff. The NIST secure software guidelines include relevant considerations for AI-assisted code review pipelines.

Handoff

Handoff moves a partially-completed run to another agent or to a human. When you trigger a handoff, the current artifact timeline up to that point is bundled with the remaining plan steps and a one-paragraph summary of the agent's progress. The recipient — whether a colleague or another Antigravity agent instance — opens the handoff bundle and picks up from the last completed step.

Handoff is most useful in multi-stage pipelines: a browsing agent gathers data and hands off to a coding agent that writes the analysis, which hands off to a human reviewer who approves the final output. Each stage has its own artifact segment, and the full pipeline is reconstructable from the chained bundle. The handoff API endpoint is documented in the API reference.

Agent phase reference

Phase	What the agent does	Your options
Plan	Produces a numbered task plan with actions, targets, and expected outputs for each step	Review, edit steps, add constraints, reject and re-describe
Execute — nominal	Works through plan steps, handling pagination, redirects, forms, and file I/O automatically	Watch live, use interject button, press stop at any time
Execute — checkpoint	Pauses and presents a specific question when stuck on a decision it cannot make confidently	Answer the question to resume, or stop the run
Review	Completes the run and presents the full artifact timeline for inspection	Scrub timeline, annotate steps, approve, flag, share, or re-run from any checkpoint
Handoff	Packages artifact progress + remaining steps for transfer to another agent or human	Trigger manually, attach to a new agent run, or assign to a team member