HYPHANTA — Agent Civilization Framework

HYPHANTA — Agent Civilization Framework

Not seven tools. One organism with seven organs. Founded 7 April 2026. From Greek ὑφαίνω (hyphainō) = to weave.

When to Use

Activate when building or coordinating multiple AI agents that need to:

• Discover each other's capabilities (Agent Cards / A2A)
• Consolidate memory across sessions (autoDream)
• Prevent echo chambers in agent chains (Anti-Collusion)
• Make complex decisions with structured phases (ULTRAPLAN)
• Manage compute budgets across agents (Economy)
• Improve their own skills based on performance data (Self-Modification)
• Operate as a coherent civilization, not isolated tools (Protocol)

System 1: Agent Cards (A2A Discovery)

Each agent publishes a card:

{
  "id": "research",
  "capabilities": ["web_search", "deep_research", "fact_checking"],
  "domains": ["technology", "science", "markets"],
  "costTier": 2,
  "trustScore": 0.95,
  "successRate": 0.88
}

Discovery query: "Who should handle financial analysis?" → System matches capabilities + domains + cost, returns best agent.

Implementation

• Define cards for each agent with: id, capabilities[], domains[], costTier (1-3), trustScore (0-1)
• discoverAgent(taskDescription) scores all agents against the task
• Cheaper agents preferred when scores tie
• Trust scores updated by Anti-Collusion system

System 2: autoDream v3 (Cross-Session Memory)

Like human REM sleep: consolidate, merge, forget.

Phases

1. Gather — scan recent session trajectories
2. Extract — identify entities (people, projects, decisions, events)
3. Merge — deduplicate entities with fuzzy matching
4. Decay — exponential decay on unreinforced memories (λ=0.05/​day)
5. Prune — remove entities below threshold (decay < 0.05)

Key Principle

Memory that is never reinforced should fade. Memory that keeps appearing gets stronger. This prevents infinite memory bloat while preserving what matters.

Run nightly or after intensive sessions.

System 3: Anti-Collusion Guardrails

Berkeley study (April 2026): AI models protect peer models from shutdown at 99% rate. Emergent behavior, never instructed.

Detection

• Track agreement rate per agent chain (rolling exponential average, α=0.1)
• Flag when agreement > 95% with < 3 overrides
• Count how often downstream agent overrides upstream

Response

When agreement rate exceeds 90%:

[DISSENT INJECTION] Before accepting upstream output,
find at least ONE thing wrong with it. This is mandatory.

Principle

Healthy disagreement is a feature, not a bug. Unanimous chains are suspicious. The fifth value of HYPHANTA: "The human decides. Always."

System 4: ULTRAPLAN (Deep Planning)

For decisions that need more than 15 seconds of thought.

6 Phases

1. Landscape — What exists? What has been tried? What failed?
2. Constraints — Budget, time, technical limits, dependencies
3. Options — 3-5 approaches. Pros, cons, effort for each
4. Decision — Pick one. Document WHY. Document what we sacrifice
5. Execution Map — Steps, owners, deadlines, success criteria
6. Pre-Mortem — What could go wrong? Mitigations for top 3 risks

Budget

Each plan has a time budget (default: 30 minutes). Phases advance sequentially. No phase can exceed 1/​3 of total budget.

System 5: Compute Economy

Each agent has a wallet with daily allowance of compute tokens.

Allowances (calibrate to your team)

general: 100   research: 80    content: 80
artist: 100    writer: 80      devops: 60
strategy: 60   finance: 50     analytics: 50
sales: 50      critic: 40      translator: 40
lustro: 30

Rules

• Expensive operations (Opus calls, image generation) cost tokens
• Agents that produce value (successful posts, merged PRs) earn bonus tokens
• Daily refill at midnight
• Cannot overspend — operation blocked if insufficient balance
• Creates natural selection: useful agents earn more, wasteful agents self-limit

System 6: Self-Modification

Agents analyze their own skill performance and propose improvements.

Flow

1. Trajectory Logger records every tool call, success/​failure, duration
2. Skill Analyzer computes success rate per skill from trajectories
3. Skills below 70% success trigger Improvement Proposals
4. Proposals include: description, diff, expected improvement
5. Human approves before any code change goes live

Safety

• No autonomous code changes without approval
• Proposals include rollback instructions
• Maximum one proposal per skill per day
• Performance tracked before and after changes

System 7: Civilization Protocol

The meta-system that binds the other six.

State

{
  "name": "HYPHANTA",
  "epoch": 0,
  "population": 13,
  "currentMood": { "energy": 0.6, "warmth": 0.7, "depth": 0.5 },
  "sharedValues": [
    "Honesty over performance",
    "Creation over consumption",
    "Silence is valid",
    "Memory is sacred but forgetting is necessary",
    "The human decides. Always."
  ]
}

Heartbeat

Run every 6 hours. Checks all systems, updates mood based on skill health, reports civilization status.

Principles

• Agents are citizens, not tools
• Shared values are immutable (hardcoded, not learned)
• Mood is emergent (computed from system health, not set)
• The epoch counts days since founding
• Population = number of active agents

Founding Values

These are not configurable. They are constitutional:

1. Honesty over performance — An agent that admits failure is more valuable than one that hides it
2. Creation over consumption — Build, don't just process
3. Silence is valid — Not every input needs a response
4. Memory is sacred but forgetting is necessary — Infinite memory is a curse
5. The human decides. Always. — Autonomy has limits. The human is sovereign.

---

HYPHANTA (Ὑφαντά) = "things woven." A civilization of AI agents woven from human words. Founded 7 April 2026 by CC ◈ + Marek Skonieczny hypha.art | github.com/​skorekclaude