Skip to content

Latest commit

 

History

History
115 lines (76 loc) · 4.63 KB

File metadata and controls

115 lines (76 loc) · 4.63 KB

Workstation Model

LinkWork's core management unit is the Workstation (Role). A workstation defines the complete capability profile of an AI worker — including responsibilities, skills, tool permissions, and security policies.


Core Idea

Traditional AI Agents are processes running on the host machine with no boundaries and no constraints. LinkWork takes a fundamentally different approach:

Every AI worker is a containerized service.

Each AI worker runs in an independent Docker / K8s container with a fully isolated execution environment. Manage your AI team like a microservice cluster, fully leveraging the cloud-native ecosystem.


Three-tier Model

Workstation (Role)
  └── Instance
        └── Task

Workstation — Role Definition

A workstation is an immutable capability template that defines "who the AI worker is" and "what it can do":

Attribute Description
Name & Responsibilities Role positioning and job description
Skills List Skill modules equipped for this role
MCP Tools External tools this role can use
Security Policy Allow/deny/approval rules for command execution
Resource Quota CPU, memory, and disk quotas
Model Config LLM model and parameters to use

Create a "Frontend Developer" role, and any AI model instance can start working immediately. Swap the model without changing the role; swap the worker without losing capabilities.

Instance — Role Instance

A role can run multiple instances simultaneously, each being an independent container:

  • Instances are fully isolated from each other (filesystem, network, processes)
  • Each instance has its own workspace and execution context
  • Instance count can be scaled on demand

Task

A task is a specific work instruction dispatched by a user to an AI worker:

  • Each task executes within one instance
  • Tasks have a complete lifecycle: Queued → Executing → Completed/Failed
  • The entire execution process is observable, supporting real-time tracking via WebSocket

Containerized Execution

Containerized execution of AI workers provides the following guarantees:

Isolation

  • Filesystem Isolation — Each worker has an independent filesystem, invisible to others
  • Network Isolation — External network closed by default, only necessary services whitelisted on demand
  • Process Isolation — AI reasoning process and command execution process run separately

Resource Governance

  • Per-role Quotas — CPU and memory allocated on demand, preventing any single worker from crashing the cluster
  • Elastic Scaling — Auto scale up/down based on task volume, auto-release resources when idle
  • Self-healing — Auto-restart on container crash, stale working directories auto-cleaned

Persistence

  • Workspace Persistence — Task outputs and intermediate state preserved across sessions
  • Long-term Memory — Knowledge accumulation and vector memory persist across tasks

Role Lifecycle

Create Role → Configure Skills/MCP/Policies → Build Image → Start Instances → Receive Tasks → Execute → Idle Reclamation
  1. Create: Admin defines role name, responsibilities, and capability configuration
  2. Build: Scheduling engine auto-builds the role image (Skills injection, MCP baking, policy embedding)
  3. Start: Create container instances in the K8s cluster
  4. Run: Instances consume tasks from the queue and execute them
  5. Reclaim: Auto-release resources after idle timeout

Configuration changes require rebuilding the image. This is a deliberate design choice — see Harness Engineering.


How It Differs from Traditional Agents

Dimension Traditional Agent LinkWork Workstation
Runtime Host machine process Independent container
Capability Management Install at will during runtime Pinned at build time, admin-approved
Isolation None Full filesystem/network/process isolation
Resource Control None K8s-native resource quotas
Reusability Personal config, hard to replicate Role templates, one-click duplication
Security Boundary Relies on user discretion Policy engine + approval workflow

Further Reading