Workspace

If you run agents on real work, you’ve likely seen these failure modes:

• a flow that worked yesterday starts behaving differently today
• personal experiment context leaks into production tasks
• identical prompts give unstable results across projects

Workspace is the first control lever that fixes this: state evolves inside a clear boundary.

What you get in practice

• stability: task context is not polluted by unrelated projects
• safety: permissions, credentials, and automations stay scoped
• recoverability: session and memory state persist per workspace

One-line mental model

Treat Workspace as an “agent partition”:

• it owns its own sessions, sources, memory, automations, and policies
• boundaries are strict by default, with no silent cross-merge

When to create a new workspace

• when objective domain changes (personal build -> team delivery)
• when risk tier changes (sandbox -> production)
• when collaborator/credential topology changes

Avoid mixing exploration + production mutation + multi-project execution in one workspace.

Minimal defaults

{
  "defaults": {
    "permissionMode": "safe",
    "workingDirectory": "/abs/path/to/repo",
    "defaultLlmConnection": "openai-prod"
  }
}

Lock these three first, then tune the rest.

Storage layout

~/.tentarc/workspaces/{workspaceId}/
├── config.json
├── sessions/
├── sources/
├── skills/
├── memory/
└── automations/

Some folders (such as memory/ and automations/) are created lazily on first use.