Questions

[Pegasus-pure]

Summary

I've built several extensions on top of Agentopia and would like to propose them for discussion — hopefully some could inform the project's future roadmap.

---

Extensions

1. Player as the 101st Agent (Implemented)

The player joins as a normal NPC. The only difference: the input source is replaced from LLM API to terminal stdin via dependency injection — the core scheduler is not modified.

• Scheduler calls an abstract input interface (LLM for NPCs, stdin for player)
• Player reuses existing event types: Solo, Joint, Public
• Simulation runs asynchronously; only pauses when player input is required

Value: Introduces a "human variable" into the agent society without architectural changes.

---

2. Time Granularity: 5×5 Phase Plan (Implemented)

Motivation: The original weekly_diary produces only 5 records/week, too coarse for observing micro-emergent behaviors.

Changes:

• PLAN stage: Agents generate a 5 days × 5 phases plan (25 items) in one shot
• ACTIVITY stage: Each day executes 5 phases, mapped to plan items
• REVIEW stage: 25 records/week instead of 5 (same logic, higher density)

Memory: Long-term preserved; short-term context reads latest 25 records (~1 week).

Value: Micro-emergence becomes observable and meaningful.

---

3. Encounter System Refactor (Implemented, feedback logic WIP)

The original system-driven Encounter event (Stage 4) is replaced with a state-driven approach:

• Trigger: Two agents both Solo + same location → auto Contact
• Emergence: Conversation outcome feeds back into the original plan — plans are mutable

The plan-feedback logic is implemented; tuning the degree of plan change via a Reward-style mechanism is planned.

Value: Enables "a conversation can change your trajectory" as a genuine emergent behavior, not a scripted event.

---

4. Dynamic Dialogue Termination (Implemented)

Problem: Fixed 20-turn dialogue is too expensive and unnatural under diary-mode density (25 events/week).

Solution:

1. Compress max turns to 8 (60% cost reduction)
2. LLM detects termination per turn: transition words ("but", "never mind") + topic drift → natural end; farewell words → early exit
3. Real-time confidence update on early termination → affects next planning cycle immediately

Value: Dialogue feels natural ("end when there's nothing left to say"), not mechanical.

---

5. Interest-Based Confidence Adjustment (Planned)

Inspired by the Reward mechanism:

• Topic-personality matchmaking factors into confidence adjustment
• Gradient decay prevents abrupt personality flips

Currently in design phase.

---

Architectural Notes

All extensions are non-intrusive:

• No core scheduler rewrite
• Dependency injection pattern for player input
• Original event types reused wherever possible
• Backward compatible with the original 8-stage weekly design

---

Would love to hear whether any of these directions align with the project's roadmap. Happy to provide more implementation details or open a PR if there's interest.