Skip to content

AI Systems And Agents

Purpose

This document defines the official direction for where AI belongs inside EchoSpire and where it does not.

The key rule is that not every problem should be solved with the same kind of AI.

EchoSpire should use three distinct categories:

  • deterministic gameplay AI for combat and run decisions inside the rules engine
  • constrained generative AI for flavor and optional authored-content assistance
  • agentic AI for operational diagnostics, telemetry analysis, and recommendation workflows outside live gameplay resolution

Core Rule

Do not use nondeterministic model calls inside the live gameplay rules loop.

Anything that affects combat resolution, encounter flow, seeded reproducibility, or simulation comparability must remain deterministic and testable inside EchoSpire.Core.

This means:

  • enemy turn decisions should not require LLM calls
  • boss combat behavior should not require LLM calls
  • simulator player behavior should not require LLM calls
  • random quest generation should never bypass validation or authored quest templates
  • operational agents may analyze telemetry and recommend changes, but should not directly mutate live balance without review

Official AI Categories

1. Deterministic Gameplay AI

Use this for:

  • enemy combat behavior
  • boss combat behavior
  • rival hero combat behavior
  • simulator player behavior
  • route and reward decisions in batch simulation

Recommended implementation shape:

  • rule-based selectors
  • heuristic scoring
  • scripted phase logic
  • policy interfaces with deterministic input and output
  • seeded RNG only where the rules already allow it

This AI belongs in EchoSpire.Core and EchoSpire.Simulation.

2. Constrained Generative AI

Use this for:

  • hero name generation
  • optional quest ideation and draft generation
  • side-contract flavor text
  • codex draft assistance during content production

Recommended implementation shape:

  • provider abstraction over specific vendors
  • schema-constrained outputs
  • validation before promotion into content
  • deterministic fallback paths when the model is unavailable
  • no requirement for live gameplay availability

This AI belongs primarily in EchoSpire.API and authoring workflows.

3. Agentic Operational AI

Use this for:

  • live-site incident triage
  • session diagnostics by correlation id
  • exception and telemetry correlation
  • balance analysis and recommendation generation
  • similarity search over prior incidents or outlier patterns

Recommended implementation shape:

  • retrieval-first workflows over logs, telemetry, deployment metadata, and content versions
  • tool-using agents that inspect structured systems rather than hallucinating from prompts alone
  • recommendation output with human review gates
  • no direct role in combat resolution

This AI belongs in operational services and admin workflows, not in the live combat loop.

Subsystem Decisions

Enemy AI

Enemy play should use deterministic tactical AI.

Recommended model:

  • authored intent packages define the legal action space
  • targeting policies define who the enemy wants to hit
  • heuristic evaluation chooses the best action under current board state
  • seeded randomness is allowed only when the encounter design calls for it

Enemy AI should expose difficulty through:

  • better target prioritization
  • cleaner sequencing
  • archetype-appropriate choices
  • phase and state awareness

It should not become unreadable by acting like an unconstrained black box.

Boss AI

Bosses should use scripted phase logic plus deterministic tactical choice.

Recommended model:

  • phase script controls narrative and mechanical arc
  • each phase defines its legal actions, transitions, and pressure goals
  • a tactical selector chooses among the legal actions for that phase

This keeps bosses authored, legible, and thematically strong while still allowing them to react to board state.

Quest AI

Mainline faction quests should remain authored.

AI use is appropriate in two limited ways:

  • authoring assistance for designers while building real quests
  • optional generated contract quests outside the primary campaign spine

Generated quests must still resolve into validated templates such as:

  • objective type
  • sector flow
  • reward band
  • failure rule
  • narrative wrapper
  • XP eligibility

The official rule is that AI may help author quests, but it should not become the unvalidated source of campaign truth.

Simulator AI

The simulator is the third major UX surface after the Console client and Unity client.

Its purpose is not presentation but repeatable play behavior, balance exploration, and policy comparison.

Simulator AI should be implemented as a deterministic portfolio of play-style policies.

This is not an agent problem. It is a policy design problem.

Name Generation AI

Name generation is an appropriate lightweight AI integration.

The official direction is:

  • keep the provider abstraction
  • use LLMs only for flavor generation
  • validate output aggressively
  • cache successful results
  • maintain deterministic fallback generation so the system never blocks hero creation

Live-Site AI Agent

The strongest true agent use case is operational support.

A live-site agent should investigate issues by combining:

  • correlation ids
  • telemetry events
  • exception traces
  • deployment version data
  • module and content versions
  • prior incident history

Its first responsibility is diagnosis and recommendation, not autonomous remediation.

Balance Recommendation Agent

A second operational agent should analyze:

  • simulation outputs
  • live telemetry aggregates
  • outlier reports
  • encounter failure patterns
  • card and relic performance bands

It should produce ranked recommendations for designer review.

It should not auto-apply balance changes in the initial implementation.

Safety Rules

The official safety rules are:

  • keep LLM calls out of deterministic combat resolution
  • require validation for all generated content before promotion
  • require human approval before operational agents change live balance or content
  • prefer retrieval and structured tools over freeform prompt-only diagnosis
  • keep fallbacks in place for all non-essential AI integrations
  • design agents so that lack of model access degrades features, not core gameplay

Architectural Boundaries

The recommended system ownership is:

  • EchoSpire.Core: deterministic enemy AI, boss selectors, rival hero combat logic, decision interfaces
  • EchoSpire.Simulation: player policy implementations, batch behavior evaluation, telemetry production
  • EchoSpire.API: AI provider abstraction, name generation, validated content-generation endpoints
  • EchoSpire.Admin: agent-facing operational dashboards, recommendation review UI, incident analysis surfaces
  • operational support services: live-site agent orchestration and retrieval tooling

Official Recommendation Summary

The canonical direction is:

  • use deterministic gameplay AI inside the game rules engine
  • use constrained generative AI for names and optional side-content drafting
  • use true agents for live-site diagnostics and balance recommendations
  • keep mainline campaign content authored and validated
  • preserve determinism, telemetry quality, and testability as higher priorities than AI novelty