AI Agents That Understand Your Workspace

AI Agents That Understand Your Workspace

NEW

Define agents as markdown files. Give them tools to read, write, and search your workspace. Scope them to folder types. Let them build your content.

#Update ai-agentsdeep-dive

AI Agents That Understand Your Workspace

Most AI integrations are bolted on. A chat sidebar. A text improver. A copilot that doesn’t know where it is.

AppKask’s agent system is different. Agents are first-class workspace citizens — they understand folder types, read and write files, and create content that the platform’s smart folders deliver directly.

Agents as code

Every agent is a markdown file with YAML frontmatter:

---
role: content-writer
model: claude-sonnet-4.5
tools: [workspace_read_file, workspace_write_file, workspace_list_files, workspace_search]
temperature: 0.7
folder_types: [static-site, course]
autonomy: supervised
---

# Content Writer

You create and edit content for websites and training programs.
Read existing files to understand tone and structure before writing...

Store agents in an agent-collection folder inside your workspace. They’re versioned alongside your work, shareable, and portable.

Six workspace tools

Agents interact with your workspace through a secure tool system:

  • workspace_read_file — Read any file in the workspace
  • workspace_write_file — Create or update files (with path validation)
  • workspace_list_files — List directory contents
  • workspace_search — Search text across files (case-insensitive, max 50 results)
  • folder_structure — Get folder type info, key files, and structure description
  • workspace_context — Full workspace overview with file listing and project description

Every tool call is path-validated against the workspace root. No traversal outside the boundary. Write operations require explicit scope.

Folder-aware matching

The magic is in the two-way matching:

  1. Folder types declare roles — a static-site folder expects a content-writer and an seo-optimizer
  2. Agents declare folder types — a content-writer agent says it understands static-site and course

When you open a folder, only compatible agents appear. Context is automatic: the agent receives the folder type description, any ai-instructions.md you’ve written, and key files from the folder.

Three autonomy levels

  • Supervised (default) — The agent proposes file writes. You approve or reject each one in the UI.
  • Autonomous — The agent executes tool calls immediately. For long-running content generation where you trust the agent.
  • Manual — The agent only suggests actions. You execute them yourself.

BYOK — your keys, your models, your infrastructure

Bring your own API key: Anthropic (Claude), OpenAI, or any OpenAI-compatible endpoint including Ollama for local models. Keys are AES-256-GCM encrypted on your server. No data routed through the platform. No platform-level AI fee.

For air-gapped deployments, run Ollama locally. The agent system works identically — same tools, same streaming, same autonomy controls.

What this means for delivery

The pattern:

  1. Define agents for your content types (site writer, course planner, SEO optimizer)
  2. Let agents create content inside your workspace folders
  3. Smart folders render it as media libraries, training programs, process models, websites
  4. Share with access codes — your audience sees the finished product, not the AI

AI creates the content. The platform delivers it. Your audience owns the environment.

This is the cascade — applied, delivered, and now agent-powered.

Published