Skip to content

ROADMAP.md

Latest commit

 

History

History
342 lines (261 loc) · 10.9 KB

ROADMAP.md

File metadata and controls

342 lines (261 loc) · 10.9 KB

Development Roadmap

A prioritized roadmap for ARGent development. Features are grouped by milestone and ordered by dependency.

Why These Milestones Matter

Every milestone serves the core vision: agents that feel alive. Here's how each phase contributes:

Milestone Contribution to Realism
1. MVP Foundation — persistent entities with memory and goals
2. Immersion Agents feel like real people — consistent, mortal, aware of time
3. Multi-Agent Agents exist beyond the player — they talk, scheme, betray
4. World Agents respond to an evolving world, not just to you
5. Memory True long-term persistence — agents remember across weeks
6. Inheritance Agent mortality and continuity — same channel, different voice

Milestone 1: Core Gameplay Loop (MVP)

Goal: Foundation for persistent entities with memory and goals

1.1 Complete Agent Response Pipeline

The current pipeline generates responses but doesn't extract state changes.

  • Trust extraction from responses

    • Classify agent responses for trust-relevant content
    • Update trust scores based on player agreement/disagreement
    • Log trust events with reasons
    • Files: src/services/classification.py, src/services/trust.py

  • Knowledge extraction

    • Extract facts revealed in conversations
    • Store in player_knowledge table
    • Inject known facts into agent prompts (avoid repetition)
    • Files: src/services/knowledge.py, src/api/inbox.py

  • Response classification

    • Classify player messages for intent, sentiment, topics
    • Store classification results
    • Use for trust calculation and trigger evaluation
    • Files: src/services/classification.py

1.2 Evidence Dashboard

The in-fiction page where players use their key.

  • Dashboard page

    • Key-derived URL: /access/<key> — the URL IS the key
    • Corporate intranet styling (90s/2000s feel)
    • Key validation against player_keys table
    • Puzzle discovery: agents hint, player figures out URL
    • Files: src/argent/api/evidence.py, src/argent/templates/evidence.html

  • Access logging

    • Log every access attempt (success/failure)
    • Track access count against limit (5 uses)
    • Emit key_used event for story triggers
    • Files: src/argent/services/evidence.py, src/argent/scheduler/events.py

  • Dashboard content

    • Security access log with timestamps
    • 2 internal memos (Project Threshold)
    • Redacted incident report
    • Email thread fragment
    • Files: src/argent/templates/evidence.html, src/argent/static/evidence.css

  • Ember visibility

    • Ember's prompt includes betrayal context when dashboard accessed
    • Knowledge fact stored: "Player accessed the evidence dashboard"
    • Agent personas include key hint reactions
    • Files: src/argent/story/prompt_builder.py, src/argent/story/personas/

1.3 Condition-Based Triggers

Currently only time-based triggers. Need condition evaluation.

  • Trigger evaluator

    • Evaluate conditions: trust thresholds, milestones, days elapsed
    • Support compound conditions (AND/OR)
    • Files: src/scheduler/evaluator.py (new)

  • Dashboard reveal trigger

    • Fire when: day >= 3 AND trust > 40 with any agent
    • Higher-trust agent reveals access
    • Files: src/scheduler/events.py

  • Post-Thursday triggers

    • Detect when Thursday deadline passes
    • Shift agent behavior (Ember desperate, Miro strategic)
    • Files: src/scheduler/events.py

Milestone 2: Immersion Systems

Goal: Agents feel like real people — consistent, mortal, aware of time

2.1 Agent Consistency (Claims Tracking)

Prevent agents from contradicting themselves.

  • Claims extraction

    • Extract significant claims from agent responses
    • Categorize: action, fact, opinion, promise
    • Mark ground truth (is this claim true?)
    • Files: src/services/claims.py (new)

  • Claims storage

    • agent_claims table with player_id, agent_id, claim, is_true
    • Index recent claims for prompt injection
    • Files: src/models/claims.py (new)

  • Consistency constraints

    • Inject recent claims into agent prompts
    • "Do not contradict: [list of previous claims]"
    • Files: src/agents/prompt_builder.py

2.2 Agent Lifecycle

Characters can fade, go silent, or disappear.

  • Lifecycle state model

    • States: engaged, cooling, silent, gone
    • Store per player-agent pair
    • Files: src/models/lifecycle.py (new)

  • Lifecycle transitions

    • Define triggers: trust thresholds, goal achieved, days elapsed
    • Transition signals (agents warn before going silent)
    • Files: src/services/lifecycle.py (new)

  • Behavioral modifiers

    • Response probability per state (1.0 → 0.8 → 0.2 → 0)
    • Response delay multiplier
    • Tone modifier (normal → terse → cold)
    • Files: src/agents/base.py

2.3 Player Engagement Tracking

Adapt pacing to player activity.

  • Engagement state model

    • States: active, casual, dormant, lapsed, churned
    • Track last message timestamps
    • Files: src/models/engagement.py (new)

  • State transitions

    • Scheduled job to update states based on inactivity
    • active (< 48h) → casual (3-5d) → dormant (7-13d) → lapsed (14-20d) → churned (21d+)
    • Files: src/scheduler/jobs.py

  • Re-engagement messaging

    • Concerned check-ins for casual players
    • Story hooks for dormant players
    • Files: src/scheduler/events.py

Milestone 3: Multi-Agent Orchestration

Goal: Agents exist beyond the player — they talk, scheme, betray

3.1 Inter-Agent Awareness

Agents learn about each other through player actions.

  • Introduction detection

    • Detect when player mentions one agent to another
    • Detect when player forwards messages
    • Files: src/services/classification.py

  • Awareness state

    • Track which agents know about each other per player
    • Store in inter_agent_exchanges table
    • Files: src/models/exchanges.py (new)

  • Knowledge sharing evaluation

    • Per-agent sharing rules (always share, always withhold, trade)
    • Deterministic evaluation based on personality
    • Files: src/services/sharing.py (new)

3.2 Coordinated Messaging

Prevent message overlap, ensure coherent pacing.

  • Message queue coordination

    • Maximum messages per day per player
    • Minimum gap between different agents
    • Files: src/scheduler/coordinator.py (new)

  • Conflict detection

    • Don't send conflicting info at the same time
    • Space out reveals
    • Files: src/scheduler/coordinator.py

3.3 Additional Agents

Expand the cast based on story progression.

  • Cipher agent

    • The intended recipient
    • Cold, technical personality
    • Spawn condition: exposure >= 50 OR Ember mentions them
    • Files: src/agents/cipher/ (new)

  • Kessler agent

    • Corporate fixer
    • Polite, professional, unsettling
    • Spawn condition: high exposure OR player causes visibility
    • Files: src/agents/kessler/ (new)

Milestone 4: World Simulation

Goal: Agents respond to an evolving world, not just to you

4.1 Exposure System

Track how "visible" the player has become.

  • Exposure tracking

    • Exposure level 0-100 per player
    • Events that increase/decrease exposure
    • Files: src/models/world.py (new)

  • Exposure triggers

    • Key used: +30
    • Dashboard accessed 3+ times: +15
    • Forwarded agent message: +20
    • Player went silent 7d: -10
    • Files: src/services/exposure.py (new)

4.2 Character Spawning

New characters appear based on player actions.

  • Spawn condition evaluation

    • Check conditions after exposure changes
    • Support: exposure threshold, agent lifecycle state, milestones
    • Files: src/scheduler/spawning.py (new)

  • Spawn messaging

    • First contact for new agents
    • Contextual introduction based on trigger
    • Files: src/scheduler/events.py

4.3 Deadline Consequences

Thursday deadline has soft consequences.

  • Thursday tracking

    • Track when 4 days have passed since game start
    • Flag thursday_passed in world state
    • Files: src/models/world.py

  • Post-Thursday behavior

    • Ember: more desperate, fewer options
    • Miro: reveals deadline was artificial
    • Exposure +20
    • Files: src/agents/*/persona.py

Milestone 5: Long-Term Memory

Goal: True persistence — agents remember across weeks and months

5.1 Memory Bank Integration

Connect to Vertex AI Memory Bank (or alternative).

  • Memory service

    • Store conversation content in Memory Bank
    • Per-player isolation
    • Files: src/services/memory.py (new)

  • Session sync

    • Save sessions to Memory Bank after key moments
    • Index claims and significant events
    • Files: src/agents/base.py

5.2 Semantic Retrieval

Pull relevant context from past conversations.

  • Memory search

    • Query Memory Bank for relevant past interactions
    • Filter by agent, topic, significance
    • Files: src/services/memory.py

  • Context assembly

    • Include relevant memories in agent prompts
    • Budget: ~400 tokens for memories
    • Files: src/agents/prompt_builder.py

Milestone 6: Channel Inheritance

Goal: Agent mortality and continuity — same channel, different voice

  • Channel ownership model

    • Track current owner per channel per player
    • Store previous owners and inheritance type
    • Files: src/models/channels.py (new)

  • Inheritance mechanics

    • Full history, partial (summaries only), or none
    • Affects what new agent can reference
    • Files: src/services/inheritance.py (new)

  • Takeover messaging

    • First occurrence: explicit announcement
    • "This isn't [previous]. I found this device."
    • Files: src/scheduler/events.py

Backlog (Future Considerations)

Quality of Life

  • Player dashboard for tracking story progress
  • "Catch me up" summary after long absence
  • Quiet hours configuration
  • Timezone-aware scheduling

Content Expansion

  • Additional story threads beyond the key
  • Modular story packs
  • Seasonal events

Multiplayer

  • Shared world state across players
  • Cross-player ripple effects
  • Collaborative discoveries

Voice Channel

  • Voicemail drops from agents
  • Voice synthesis for agent calls
  • Speech-to-text for player responses

How to Use This Roadmap

  1. Pick a milestone — Work through milestones in order; they build on each other
  2. Check dependencies — Some features require others (e.g., triggers need classification)
  3. Mark complete — Check off items as they're implemented
  4. Update as needed — This is a living document; adjust based on learnings

Priority Guidelines

  • Milestone 1 is required for a playable MVP
  • Milestone 2 makes agents feel real
  • Milestone 3 enables the multi-agent experience
  • Milestones 4-6 add depth but aren't blocking