💡 Agent Sandbox Mode: Risk-Free Exploration Environment for Non-Technical User Onboarding

[github-actions[bot]]

Summary

A dedicated sandbox mode where agent actions are simulated rather than executed, allowing non-technical users to explore TalkTerm's capabilities without fear of unintended consequences. The avatar narrates what it WOULD do, files appear in a temporary workspace, and external calls are mocked. Users switch to live mode when confident. This directly complements the Guided Interactive Onboarding (Discussion #245, approved and in active development via issues #333-338).

Market Signal

Gallup research shows non-technical workers' #1 barrier to AI adoption is inability to see how AI fits their specific tasks. WRITER's 2026 survey found 37% cite limited time to learn and 30% cite difficulty finding reliable information as top obstacles. A sandbox eliminates the cost of experimentation, letting users discover use cases organically without risk. No competing desktop agent (ChatGPT Voice Desktop, Hermes Agent, Claude Code) offers a sandbox exploration mode.

User Signal

• TalkTerm's onboarding epic (issues Guided Interactive Onboarding: avatar-led learn-by-doing first session #333-338) is actively being built — the tutorial runs through real workflows, but a first-time user saying "yes" to file creation or API calls during a tutorial carries real consequences
• 79% of enterprises face AI adoption challenges (WRITER 2026) — fear of unintended consequences is a significant barrier
• The confirm-plan gate (FR20) helps but interrupts flow; sandbox mode removes the gate entirely by making ALL actions consequence-free

Technical Opportunity

Feasible with the existing architecture:

• Create a sandboxed temp workspace using the WorkspacePath branded type
• Configure AgentBackend with a dry-run flag or mock tool execution via the existing abstraction layer
• Add a SandboxIndicator UI component to StatusIndicator (UX-DR6) — e.g., a warm-colored "Sandbox" badge
• The STT/TTS abstraction layer and overlay system work identically in both modes — only the agent execution path changes
• SessionMode enum (live | sandbox) in src/shared/types/domain/

Assessment

Dimension	Score	Rationale
Feasibility	high	Temp workspace + dry-run execution are well-understood patterns; architecture already supports swappable backends
Impact	high	Directly addresses the #1 barrier to non-technical user adoption; makes onboarding tutorial zero-risk
Urgency	med	Should be ready before onboarding epic completes; not blocked by external factors

Adversarial Review

Strongest objection: TalkTerm already has confirm-plan (FR20) for every destructive action. Sandbox mode is redundant with the existing safety mechanism.

Rebuttal: Confirm-plan is a gate that interrupts workflow ("do you want to do this?"). Sandbox mode is a mindset ("nothing I do here has consequences"). They serve different psychological needs. For onboarding, sandbox mode eliminates the anxiety of accidentally approving something harmful. The user explores freely, builds confidence, then switches to live mode. Confirm-plan remains active in live mode as the production safety layer.

Suggested Next Step

Define a SessionMode enum (live | sandbox) in src/shared/types/domain/. Add sandbox workspace provisioning to the storage layer. Wire the sandbox indicator into the StatusIndicator component (UX-DR6). Coordinate with the Guided Interactive Onboarding epic to use sandbox mode for tutorial sessions.