Agentic System Design
Our agentic browser automation system is made up of two cooperating agents: The builder and the maintenance agent. Together, these agents enable a robust automation lifecycle: from rapid creation to continuous adaptation. By decoupling generation from runtime reasoning, the system ensures faster execution, lower operational cost, and improved reliability compared to taditional browser-use frameworks that use LLMs at every single step at runtime.
| Agent | Responsibility |
|---|---|
| Builder Agent | Creates the initial browser‑automation graph via chat with human or AI. |
| Maintenance Agent | Monitors live runs and repairs the graph whenever the target site changes. |
Builder Agent
Purpose
Turn a goal or standard operating procedure (SOP) into a browser agent that can be triggered over our Run API.
How it works
- Kick‑off – Open a new workflow in the dashboard and describe the task. You can paste a detailed SOP or simply type a high‑level goal like “Fill out this onboarding form.”
- Schema discovery – Optionally provide JSON schemas for input variables and expected output. Skip this, and the agent infers them automatically.
- Interactive build – The agent launches a live browser preview, navigates the target site, and asks clarifying questions only when needed (e.g., login credentials, ambiguous clicks).
- Credential capture – Any secrets you share are vaulted immediately in Authentication Center and reused securely on subsequent runs.
- Graph generation – Each confirmed action converts into a node‑edge graph you can inspect, version, and reuse across environments.
- Save & run – Once the agent completed the graph generation, hit Save & Run to execute the first full, deterministic run.
Why we generate code instead of reasoning at every step
Traditional “agentic” tools keep an LLM in the loop for every click, form fill, and page transition. CloudCruise compiles the logic up front and reuses it. The payoff shows up in three dimensions:
| Dimension | Runtime LLM at every step | CloudCruise compile‑once approach |
|---|---|---|
| Speed | Each action incurs network latency plus token inference time, leading to multi‑minute runs. | Graph executes at native browser speed; only API calls hit the network. |
| Cost | Pay per token, per step—cost scales linearly with run length. | One‑time build tokens are amortized over thousands of executions; marginal cost is near‑zero. |
| Reliability | High variance: LLM may hallucinate selectors or loop indefinitely; reproducibility is low. | Fully deterministic; every run follows the same graph. Maintenance Agent handles drift proactively. |
Maintenance Agent
When a run fails, the Maintenance Agent wakes up, labels the failure, and triggers the right recovery path - notify, auto‑repair, or retry. The classifier uses the matrix below.
Error‑classification matrix
| Category | Sub‑category | Description |
|---|---|---|
| User Error | PAGE_NOT_FOUND | URL returns 404; usually bad ID/slug |
AUTHENTICATION_ERROR | Wrong or expired credentials, unexpected 2FA, captcha, password reset | |
INCORRECT_FORM_INPUTS | Provided value empty, invalid, or fails validation | |
PASSWORD_UPDATE_REQUIRED | Site forces password change before further access | |
ACCOUNT_UPDATE_REQUIRED | Site demands profile update or new T&C acceptance | |
ADDITIONAL_USER_INPUT_REQUIRED | Unexpected gating modal that truly blocks progress | |
MULTIPLE_MATCHING_RESULTS_FOUND | Ambiguous search results require human or AI disambiguation | |
ACTION_BLOCKED_BY_PLATFORM | Platform rejects duplicate or forbidden action | |
| Workflow Error | ACTION_PERFORMED_TOO_EARLY | Clicked before element was fully loaded |
UNEXPECTED_POPUP | Optional modal (survey, promo, cookie banner) appeared | |
UNEXPECTED_UI_STATE | Layout/state differs (already logged in, collapsed sidebar) | |
XPATH_INCORRECT | Selector matches 0 or >1 elements | |
| External Error | SERVICE_UNAVAILABLE | Upstream system down or non‑responsive |
Recovery playbook
| Classification | Action |
|---|---|
| User Error | Surface a dashboard alert + notification; ask the user to correct data or credentials. |
| Workflow Error | Auto‑patch the graph: update selectors or insert waits. |
| External Error | Schedule exponential back‑off retries; no graph change. |
The result is high run success without burning compute on real‑time reasoning - most fixes are either user‑driven or one‑shot graph patches rather than repeated LLM calls.