Skip to content

isolated-workspaces.md

Latest commit

 

History

History
64 lines (45 loc) · 2.69 KB

isolated-workspaces.md

File metadata and controls

64 lines (45 loc) · 2.69 KB

Isolated Workspaces

Problem

When multiple agents share a workspace, they contaminate each other's context. An engineering agent's half-finished refactor confuses the ops agent's log analysis. Client A's credentials appear in the context when working on Client B. One agent's bad session pollutes another's memory.

Symptoms:

  • Agents reference files they shouldn't know about
  • Context windows fill with irrelevant information
  • Resetting one agent's context requires resetting all of them
  • Security boundaries between clients/projects are unclear

Solution

Give each agent its own filesystem workspace with separate memory, credentials, and configuration.

workspace-engineer/
├── IDENTITY.md
├── SOUL.md
├── MEMORY.md
├── memory/
│   ├── 2026-03-25.md
│   └── 2026-03-26.md
└── projects/

workspace-ops/
├── IDENTITY.md
├── SOUL.md
├── MEMORY.md
├── memory/
│   └── 2026-03-26.md
└── runbooks/

Implementation

  1. Create a workspace directory per agent. Each agent's working directory is its own root.

  2. Scope credentials. Each agent only has access to the credentials it needs. The research agent doesn't need SSH keys. The ops agent doesn't need the design tool API key.

  3. Separate memory files. Each agent maintains its own MEMORY.md (long-term) and memory/ directory (daily logs). No shared write access to memory.

  4. Shared read-only resources. Some things should be shared (company docs, style guides). Mount or symlink these as read-only in each workspace.

  5. Coordinator consolidation. If you need a unified view, designate one agent (the orchestrator) to read from multiple workspaces and consolidate. One-way: agents don't read each other's workspaces directly.

Trade-offs

  • Disk space: Multiple copies of shared resources. Minimal in practice (text files are tiny).
  • Coordination overhead: Information doesn't flow between agents automatically. You need explicit sharing mechanisms.
  • Setup complexity: More directories to manage. Automation helps (script the workspace creation).

When to Skip

  • You have only one agent. Isolation is meaningless with a single workspace.
  • All agents work on exactly the same project with no client separation needs.
  • Your agents are stateless (no memory between sessions). Isolation matters most for persistent context.

Related Patterns