💡 Progressive Trust Escalation: Adaptive Agent Autonomy Through Usage Patterns

[github-actions[bot]]

Summary

Implement a dynamic trust system where agent autonomy increases as users demonstrate competence — starting with full confirmation for every action and gradually reducing friction based on usage history, action classification, and demonstrated reliability. Trust levels are per-action-class (read operations escalate faster than writes/deletes), user-visible, and manually adjustable.

Market Signal

Progressive delegation is now the dominant UX pattern for agent adoption in 2026 (Gartner, industry UX research). Claude Cowork uses progressive delegation. GitHub Copilot CLI shipped suggest-to-auto-apply escalation in June 2026. Every major agent platform is moving away from binary confirm/deny toward graduated trust. OpenAI's full-duplex voice mode in testing further validates natural, low-friction agent interaction as the market expectation.

User Signal

TalkTerm's PRD (FR20) defines a binary confirmation gate — confirm or deny every destructive action. Non-technical users — the primary target — will abandon the product if every file write or API call requires manual approval. Conversely, auto-approving everything creates safety risk. The confirm-plan pattern needs a trust dimension to balance safety and conversational flow. Related: FR20 (confirmation gate), FR21 (approve/modify/reject plans), FR22 (hidden intermediate thinking).

Technical Opportunity

The existing permission callback pattern in the Claude Agent SDK (allowedTools, permission callbacks) already supports action classification. TalkTerm's AgentBackend abstraction and IPC confirm channel (agent:confirm) provide the architectural seam for inserting trust logic without changing the SDK integration. Trust state persists naturally in the existing electron-store profile alongside user preferences (FR36). The SDK's maxBudgetUsd and maxTurns controls complement trust escalation with hard safety limits.

Assessment

Dimension	Score	Rationale
Feasibility	high	Uses existing SDK permission callbacks, IPC channels, and electron-store. No new infrastructure needed — pure application logic.
Impact	high	Directly addresses the adoption cliff where non-technical users churn due to confirmation fatigue or safety anxiety.
Urgency	high	Progressive delegation is now table stakes for agent UX. Should be designed into Epic 5 (Decision Presentation) from the start.

Adversarial Review

Strongest objection: If trust escalates too aggressively, users lose control over consequential actions. If too conservative, it becomes the existing confirm-plan pattern with extra complexity and no user benefit.

Rebuttal: Trust levels are action-class-specific (read operations escalate faster than writes/deletes) and user-visible (a trust indicator shows current autonomy level with milestones like 'You've approved 10 file reads — I can now read files automatically'). Users can manually adjust trust per action class at any time. The system resets trust for new workspaces or sensitive operations (API calls, deletions). This is a spectrum with guardrails, not a binary switch.

Suggested Next Step

Design the trust tier model (3-4 tiers: full-confirmation, selective-confirmation, auto-approve-reads, auto-approve-writes), action classification taxonomy (read/write/delete/external-API/shell), and escalation criteria (N successful approvals per class). Write as a technical design doc planning artifact before Epic 5 implementation.