claude-code - 💡(How to fix) Fix [FEATURE] Feature Idea — Trust Calibration and Progressive Autonomy Model [2 comments, 2 participants]

Preflight Checklist

• I have searched existing requests and this feature hasn't been requested yet
• This is a single feature request (not multiple features)

Problem Statement

I'd like to propose a trust calibration system that allows the relationship between a user and Claude to deepen over time, rather than resetting with each session.

The core idea: Claude's level of autonomy should be earned and domain-scoped, not fixed. A tiered model could allow users to progressively unlock different modes of interaction based on demonstrated alignment over time.

Key dimensions:

Session continuity — Currently Claude has no persistent identity across sessions. A trust tier system would require some memory of past interaction quality and outcomes, so that earned trust isn't lost at session end.

Domain-scoped trust — Trust shouldn't be monolithic. A user might trust Claude at full autonomy for code review and refactoring, while preferring a more cautious, confirmation-seeking mode for financial decisions or irreversible actions. Tiers should be configurable per domain.

Earned autonomy progression — The natural ladder would mirror how you'd onboard a human engineer:

1. Ask before acting — confirm every significant step
2. Act and report — proceed, then summarize what was done
3. Full autonomy — operate independently within agreed scope

Deepening relationship — The current model treats every session as first contact. This idea recognizes that user-Claude collaboration should compound, not reset — the longer the working relationship, the more calibrated and efficient it becomes.

This feels particularly relevant to Anthropic's alignment work: it's not just a UX improvement, it's a framework for making autonomous AI behavior legible and accountable as it scales up.

Proposed Solution

1. Trust Profile (Persistent, Per-User)

Store a lightweight trust profile tied to the user's account — not raw conversation history, but distilled signals:

• Domains where the user has granted elevated autonomy (e.g., code, writing, research)
• Autonomy tier per domain (Supervised / Act-and-Report / Full Autonomy)
• Optional: a brief user-written context ("I'm a senior engineer, treat me accordingly")

This profile travels across sessions and devices.

---

1. Three Autonomy Tiers

┌───────────────┬───────────────────────────────────────────────┬────────────────────────────────────────┐ │ Tier │ Behavior │ Default for │ ├───────────────┼───────────────────────────────────────────────┼────────────────────────────────────────┤ │ Supervised │ Ask before every significant action │ New users, sensitive domains │ ├───────────────┼───────────────────────────────────────────────┼────────────────────────────────────────┤ │ Act & Report │ Proceed, then summarize what was done │ Established users, moderate-risk tasks │ ├───────────────┼───────────────────────────────────────────────┼────────────────────────────────────────┤ │ Full Autonomy │ Operate within agreed scope without check-ins │ Trusted users, low-risk domains │ └───────────────┴───────────────────────────────────────────────┴────────────────────────────────────────┘

---

1. Progression Mechanism

• Users can manually set their tier per domain at any time
• Claude can suggest an upgrade when patterns indicate the user consistently approves actions without modification ("You've approved every file edit this session — want me to act and report instead?")
• Users can override downward at any time, including mid-session ("be more cautious from here")

---

1. Domain Scoping

Trust tiers are namespaced, not global. Example profile:

code_editing: full_autonomy file_deletion: supervised financial: supervised research/writing: act_and_report

This prevents a high-trust code relationship from bleeding into unrelated high-stakes domains.

---

1. Auditability

Every action taken under Act & Report or Full Autonomy gets logged to a session summary the user can review — making autonomous behavior legible and correctable, which directly supports alignment goals.

---

Why this works for Anthropic

• It doesn't require trusting Claude more by default — it lets users calibrate trust explicitly
• It creates a natural feedback loop that improves both UX and safety
• It mirrors how trust works with human collaborators, making the model more intuitive

Alternative Solutions

No response

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

No response

Additional Context

No response