File Ownership

Agents interact with your project files in controlled ways. You provide the input, choose where output goes, and the agent processes content within those boundaries.

How input works

You select what an agent processes by configuring Input Sources in the deployment. This can include:

  • Individual files from your project
  • Entire folders
  • Workspace documents

The system loads this content and injects it into the pipeline before execution begins. Agents do not browse your project looking for files to process — you explicitly choose what to give them.

You own input selection. The agent processes what you provide. It does not autonomously decide which files to read as its primary content.

Reading files

Agents can read files in two ways:

  • Input sources (primary) — the files you drag into the deployment context. These are loaded automatically before the pipeline runs.
  • Secondary reads — if the agent's pipeline includes a step that needs a template or reference file (determined by an earlier step's output), it can read that specific file. This is for supplementary data, not for fetching primary content.

Agents can also search and list files in the project (find files by name, check metadata) when their pipeline requires it.

Writing files

When an agent produces output that should be saved, it writes to the Output Folder you configured on the deployment.

Key details about file writing:

  • Files are written with a standard naming pattern: AgentName_YYYYMMDD_HHmmss.format
  • The output format (Markdown, text, JSON, CSV, HTML) is controlled in agent settings
  • You can toggle file saving on or off without rebuilding the agent
  • If no output folder is configured, the save step is skipped silently (output still displays on screen)
  • Agents create new files — they do not overwrite or modify your existing files

What agents cannot do with files

For safety, user-created agents cannot:

  • Delete files from your project
  • Move files between folders
  • Copy files to different locations
  • Overwrite existing files
  • Write outside the configured output folder
  • Access files through directory traversal or unsafe paths

These restrictions ensure that running an agent does not reorganize or destroy your existing project content.

Reference files

The Reference (LLM Rules / Directives) area in a deployment lets you attach files that inform how the agent works without being the main content to process. Examples:

  • Brand voice guidelines
  • Output templates
  • Style rules or formatting specifications
  • Domain vocabulary or terminology lists

Reference files are read-only context. The agent uses them to shape its output but does not modify them.

Practical workflow

  1. Deploy an agent to Global or Current Project scope
  2. Drag input files into the Input Sources area
  3. Set an Output Folder where results should be saved
  4. Optionally attach reference files for domain context
  5. Run — the agent reads your input, processes it through the pipeline, and writes output to your chosen folder

You can change input sources and output folder anytime without rebuilding the agent. The configuration persists between runs.

Path safety

All file paths used by agents are validated before any read or write operation. ThinkForge blocks:

  • Paths that attempt directory traversal (moving up out of the project)
  • System paths outside your project structure
  • Paths with invalid or dangerous characters

This ensures agents operate within the boundaries of your project and output folder.

Related