Agentic Workflow
Agentic Workflow describes task objectives in natural language and uses a large language model (LLM) to autonomously plan workflow steps. It can operate the browser, access files in the working folder, send emails through the email account linked to EMILY, and write code to complete the objective. Note that larger task goals are more likely to cause the language model to make incorrect judgments and will also use more tokens. We recommend breaking them down into multiple smaller tasks to reduce the failure rate. If you are not familiar with prompting a language model to achieve a task objective, you can use "Ask EMILY" to let AI generate a model-friendly objective prompt.

Parameters
API KEY - The API key for OpenAI / Google / Anthropic. Supports the %FILENAME% variable, or you can use the dedicated stored-value key %credit-key%.
- For an OpenAI API key, see Apply for an OpenAI Key
- For a Google API key, see Apply for a Gemini Key
- For an Anthropic API key, see Apply for an Anthropic Key
MODEL - The following models are currently supported:
| Platform | Model | Pricing |
|---|---|---|
| OpenAI | gtp-5.4, gpt-5.4-mini | OpenAI official website |
| gemini-3-flash, gemini-3.1-pro, gemini-3-flash | Gemini official website | |
| Anthropic | claude-opus-4-6, claude-sonnet-4-6 | Anthropic official website |
BUDGET - The task budget, priced in USD. The task will stop once it exceeds the budget.
COST - The actual task cost, priced in USD. It is updated dynamically during test runs or production runs.
SANDBOX - Sandbox settings. The Agent is already restricted to accessing only the built-in browser and the working folder. You can further restrict prohibited web pages or files in the working folder.
SKILLS - Sub-skills that can be called during the task. Each sub-skill should fully describe its function, when to use it, and its input and output files.
MISSION - Describes the task in natural language. The %FILENAME% template can be used.
Example
Using this patent search as an example, the task content should clearly describe which website to query, what conditions to use for the search, the output file format, and the objective to be achieved:

If an enterprise has special requirements for data confidentiality, click "SANDBOX" to set a blacklist of prohibited websites and working-folder files:

For deterministic operations or the handling of sensitive files, you can use RPA instead and provide it as a sub-skill (SKILL) for Agentic Workflow to call, click "SKILLS" to add:
