💡 Deliberation Mode: Avatar as Structured Thinking Partner Without Execution

[github-actions[bot]]

Summary

Add a 'thinking partner' mode where the avatar acts as a structured sounding board — asking probing questions, identifying assumptions, exploring alternatives, and playing devil's advocate — without executing any agent actions. Users toggle between deliberation (think it through) and execution (make it happen), creating a natural on-ramp from reflection to action that's especially valuable for non-technical users who may be intimidated by agent execution.

Market Signal

GitHub Copilot CLI shipped a 'rubber duck' agent in June 2026 that provides constructive criticism of plans and implementations — a structured thinking partner integrated into the agent workflow. ChatGPT's voice mode without tools already serves a deliberation function for millions of users. The pattern of separating thinking from doing is emerging as a distinct, valued product capability. Gartner's agent UX research identifies 'progressive delegation' as the dominant adoption pattern — deliberation mode is the natural first rung of that ladder.

User Signal

TalkTerm's confirm-plan pattern (FR20/FR44) and execution-oriented design create friction when the user just wants to think out loud. Consider Sarah (Journey 1 persona): she says 'I'm not sure which approach to take for our onboarding redesign.' The current model would have the agent start executing a brainstorming workflow. What Sarah actually wants is a thinking partner who helps her clarify her own thoughts before delegating execution. Non-technical users may use TalkTerm as a thinking tool before they're ready to use it as an execution tool — making deliberation mode both a feature and an adoption bridge.

Technical Opportunity

Deliberation mode requires no new infrastructure — it operates on the existing voice pipeline, avatar, overlay system, and agent backend. The implementation is a system prompt modifier that instructs the Claude agent to probe rather than act, plus a mode parameter on the AgentBackend interface (startSession(config: { mode: 'deliberation' | 'execution' })). ActionCards (UX-DR2) present structured thinking frameworks instead of execution options (e.g., 'Consider the opposite perspective', 'What would your user say?', 'List your top 3 assumptions'). The overlay system (UX-DR1) works unchanged. Cost is minimal — deliberation conversations use fewer tokens than execution workflows since no tool calls occur.

Assessment

Dimension	Score	Rationale
Feasibility	high	System prompt change + mode toggle. No new infrastructure, no architecture changes. Reuses entire existing pipeline.
Impact	med	Valuable for non-technical users as adoption on-ramp. Niche but defensible — no competitor offers structured voice-first deliberation with an avatar.
Urgency	low	Phase 2+ feature. Core execution workflows (Epics 4-6) should ship first. But the AgentBackend interface should anticipate mode as a parameter.

Adversarial Review

Strongest objection: Users might stay in deliberation mode and never execute, reducing TalkTerm to 'just another chatbot' that talks but doesn't do. The product's core differentiator is real execution capability — deliberation mode actively undermines that differentiator.

Rebuttal: Deliberation mode is a bridge, not a destination. When a plan or decision emerges from thinking, the avatar proactively suggests transitioning to execution: 'We've identified three approaches — want me to start building option A?' This creates a natural on-ramp from reflection to action. The mode switching is visible and intentional — users see they're in 'Thinking' mode and can switch to 'Doing' mode at any time. GitHub Copilot's rubber duck agent validates that structured criticism improves execution quality rather than replacing it. For TalkTerm's non-technical users, building confidence through successful deliberation makes them MORE likely to trust the agent with execution, not less.

Suggested Next Step

Design 5-7 structured thinking techniques the avatar can guide in deliberation mode: assumption mapping, pre-mortem analysis, options matrix, stakeholder perspective rotation, constraint identification, success criteria definition, and risk assessment. Prototype as a system prompt variant with the existing Claude Agent SDK integration. Test whether deliberation-to-execution transitions feel natural or disruptive.