Token Economy & Reputation

Rnix implements a token economy with budget pools, contract SLAs, agent reputation scoring, and Skill synergy emergence for optimized multi-agent resource allocation.

---

Token Budget Pools

Each Compose workflow is assigned a total token budget pool. Agents request allocations from this shared pool.

Configuration

# compose.yaml
version: "1.0"
intent: "Code review workflow"
budget_pool:
  total: 50000        # Total tokens for the entire workflow
  allocation: priority # priority | equal | proportional

agents:
  analyzer:
    intent: "Analyze code quality"
    agent: "code-analyst"
    budget: 20000      # Agent-level budget (drawn from pool)
    priority: high     # Affects allocation during contention
  doc-gen:
    intent: "Generate docs"
    budget: 15000
    priority: normal

Allocation Strategies

Strategy	Behavior
priority	High-priority and critical-path agents get more tokens; low-priority tasks are queued or degraded
equal	Even split across all agents
proportional	Based on historical consumption patterns

When agents compete for limited budget, the price signal mechanism schedules allocation — critical-path agents get priority, non-critical tasks wait.

Budget Enforcement

Per LLM call:
  proc.TokensUsed += response.TokensUsed

  if budget > 0 && TokensUsed >= budget:
    finishProcess(ExitStatus{Code: 2, Reason: "budget_exceeded"})

Exit codes: 0 = normal, 1 = error, 2 = budget exceeded.

---

Contract SLA

Agents collaborate through explicit contracts that define quality expectations:

Contract Fields

Field	Description
input_format	Expected input structure
output_quality	Quality criteria (pattern or LLM-evaluated)
max_tokens	Maximum token consumption
timeout	Maximum execution time
sla_level	bronze / silver / gold

SLA Evaluation

After contract execution completes, the system automatically evaluates:

1. Output quality — Does the result meet the defined criteria?
2. Token efficiency — Was the token budget used effectively?
3. Response time — Was the execution within timeout?

Evaluation results feed into the reputation system.

---

Agent Reputation System

Every Agent template accumulates a reputation score based on historical performance:

$ rnix reputation code-analyst
Agent: code-analyst
  Executions:    47
  Success rate:  93.6%
  Avg tokens:    2,340
  SLA met:       89.4%
  Reputation:    ★★★★☆ (4.2/5.0)

$ rnix reputation            # List all agents by reputation

Reputation Metrics

Metric	Weight	Description
Success rate	40%	Percentage of tasks completing without error
Token efficiency	25%	Average tokens vs budget utilization
SLA compliance	25%	Percentage of contracts meeting SLA
Speed	10%	Average execution time vs timeout

Auto-Selection

When the Reconciler or Compose engine needs to select an Agent template (e.g., for intent decomposition), it preferentially selects high-reputation templates — a natural selection mechanism where better-performing agents are used more.

---

Skill Synergy Emergence

When certain Skills are loaded together, they can produce emergent capabilities beyond their individual functions.

Declaring Synergy

# skills/security-scan/SKILL.md frontmatter
---
name: security-scan
description: "Scan for security vulnerabilities"
allowed-tools: /dev/fs /dev/shell
synergy:
  code-analysis:
    description: "When combined with code-analysis, enables deep security-aware code review"
    instructions: |
      With both security-scan and code-analysis active, you can:
      - Correlate code quality issues with security implications
      - Identify security anti-patterns in code structure
      - Generate security-annotated code review reports
---

Auto-Detection

When an agent loads both security-scan and code-analysis, Rnix automatically:

1. Detects the declared synergy
2. Appends synergy instructions to the system prompt
3. Records the combination's performance in the synergy matrix

Synergy Matrix

$ rnix synergy list
Known effective Skill combinations:
  security-scan + code-analysis     → +23% quality  (47 observations)
  code-analysis + dependency-check  → +15% coverage (32 observations)

The matrix tracks which Skill combinations historically produce significantly better results than individual Skills, based on reputation system data.

---

Related Documentation

• Compose Orchestration — Multi-agent workflows with budgets
• Autonomous Agents — Unified reasoning and stem cell differentiation
• Agents & Skills — Agent and Skill definitions
• Security — Adaptive immune security