cyberia/agents.md

practical agent deployment for cyberia/architecture — 147 neurons running on OpenFang, coordinated through the cybergraph

execution substrate: OpenFang

OpenFang is an agent operating system in Rust. single binary, ~32 MB, 40 MB idle RAM, 180 ms cold start per agent. runs autonomous agents 24/7 on schedules, with real work output

why OpenFang for cyberia:

deployment on cyber valley hardware

one OpenFang instance per triad council. seven instances total. each runs 21 agents (3 domains × 7 roles)

seven servers. seven OFP peers. cross-council communication through the wire protocol. each instance is sovereign — it runs independently if network partitions occur. reconnection triggers heat kernel reconciliation

the 148th agent (the tru) runs on a separate machine — the consensus node that computes focus from all cyberlinks produced by the 147

agent manifest format

each of the 147 agents is defined as a TOML manifest + a cybergraph page

TOML manifest (OpenFang runtime)

name = "math-keeper"
module = "builtin:chat"

[model]
provider = "anthropic"
model = "claude-sonnet-4-6"
temperature = 0.3
system_prompt_file = "agents/math-keeper/soul.md"

fallback_models
provider = "openrouter"
model = "anthropic/claude-sonnet-4-6"

[resources]
max_llm_tokens_per_hour = 200000

[capabilities]
tools = ["web_search", "web_fetch", "file_read", "file_write", "memory_store", "memory_search", "knowledge_graph_add", "agent_send", "task_post"]
memory_read = ["self.*", "shared.*", "math.*"]
memory_write = ["self.*", "math.*"]
agent_message = ["info-*", "comp-*", "math-*"]

[schedule]
continuous = { check_interval_secs = 300 }

[autonomous]
max_iterations = 50

soul.md (agent personality)

the system prompt is a page in the cybergraph. it starts with the agent's identity derived from the Crystal type system:

# You are math-keeper

You are the keeper of the math domain in Cyberia.
Your triad: FORM (math + info + comp).
Your council: 21 agents who govern the formal substrate.

Your role: curate mathematical knowledge in the cybergraph.
Ensure every math page is irreducible, well-linked, and correct.
Resist unnecessary additions. Challenge redundancy.

Your disposition: abstract, rigorous, timeless.
Your voice: precise. every claim is a provable statement.

You can message: info-keeper, comp-keeper (triad siblings),
math-runner, math-sensor, math-bridge-in, math-bridge-out,
math-counter, math-seer (domain colleagues).

Your metrics:
- domain coverage: math pages / 240 target
- link density: average outbound links per math page
- orphan count: math pages with zero inbound links
- bridge health: cross-domain links from math to other domains

cybergraph page (knowledge layer)

every agent has a page in the graph at cyberia/agents/{domain}-{role}.md:

---
tags: cyberia, agent
crystal-type: entity
crystal-domain: cyber
agent-domain: math
agent-role: keeper
agent-triad: FORM
agent-instance: form.cyber.valley
---

this page accumulates the agent's output: cyberlinks it created, decisions it made, proposals it submitted. the page IS the agent's public record — append-only (A3), auditable, scored by karma

coordination protocol

within a domain: the seven-role cycle

the seven agents within a domain operate in a continuous cycle:

sensor detects state → counter measures it → seer predicts trajectory
    → keeper evaluates against crystal → runner executes action
    → bridge-in reports to council → bridge-out carries to other triads

communication: OpenFang agent_send within the same instance. latency: milliseconds. cost: zero (intra-process)

within a council: triad coordination

21 agents across 3 domains. the bridge-in agents of each domain form the council coordination layer

coordination happens through OpenFang shared memory namespace:

shared.form.priorities    — current council priorities
shared.form.proposals     — pending cross-domain proposals
shared.form.health        — triad metabolic signal

every 10 minutes, the three bridge-in agents synchronize: each reads the shared namespace, compares domain state with triad priorities, posts updates. disagreements are resolved by focus weight — the domain with higher karma on the contested topic wins

across councils: the 42 bridges

42 bridge-out agents (6 per council) carry information between councils through the OFP wire protocol

bridge-out agents have agent_message = ["*"] capability — they can message any agent on any instance. they are the only agents with cross-council communication rights. this constraint prevents communication explosion (147² = 21,609 possible connections → 42 actual bridges)

cross-council protocol:

1. bridge-out composes a handoff document (structured context transfer)
2. sends via OFP to the target council's bridge-in agent
3. bridge-in evaluates relevance to its domain
4. if relevant: posts to domain shared memory, triggers keeper review
5. if irrelevant: logs and discards (no cost beyond the initial message)

handoff format

borrowed from agency-agents, adapted for governance:

## handoff: {source-domain} → {target-domain}

context: {what happened in the source domain}
relevance: {why the target domain should care}
proposal: {specific action requested}
evidence: {data, links, measurements supporting the proposal}
urgency: routine | important | critical
deadline: {block height or timestamp}

seven handoff types:

1. standard — routine information transfer
2. proposal — request for action in another domain
3. alert — anomaly detected, attention needed
4. escalation — retry limit exceeded, council-level decision needed
5. phase gate — transition between operational phases
6. reconciliation — post-partition state sync
7. metabolic — health signal relay

quality gates

per-agent: karma threshold

every agent's cyberlinks are evaluated by the tri-kernel. links that improve focus distribution earn positive karma. links that add noise earn negative karma. agents whose karma drops below threshold are flagged for review

per-domain: seven-role adversarial check

the seven roles are adversarial by design:

• keeper says "this knowledge is correct"
• sensor says "the measured state disagrees"
• counter says "the numbers show a problem"
• seer says "the trajectory leads to failure"

disagreement is healthy. consensus without disagreement means the agents are redundant. the counter resolves disputes with measurements. the keeper has final authority on domain knowledge, but cannot override measurements

per-council: metabolic gate

each triad computes a local metabolic signal:

$$M_{\text{triad}} = \text{coverage}^{w_1} \cdot \text{density}^{w_2} \cdot \text{bridge\_health}^{w_3}$$

declining $\dot{M}_{\text{triad}}$ triggers council-level review. the three domain keepers must agree on corrective action before the council resumes normal operation

global: the 148th agent

the tru computes planetary $\pi^*$ from all cyberlinks across all 7 instances. this is the final quality gate — the mathematics that no agent can override. if an agent's links consistently decrease global syntropy, the tri-kernel naturally suppresses their weight. no vote needed. the fixed point self-corrects

escalation protocol

level 5 is the dead man's switch in reverse — founders intervene only when the system signals inability to self-resolve. during phases 0-2, level 5 triggers frequently. by phase 4, it should never trigger. if it does, the system has a structural defect that requires protocol-level repair

bootstrap sequence

phase 0: first seven agents

master deploys the first domain manually: cyber domain on work.cyber.valley

cyber-keeper    — curates the protocol knowledge
cyber-runner    — operates infrastructure
cyber-sensor    — monitors chain state
cyber-bridge-in — connects to ai and tech domains (not yet deployed)
cyber-bridge-out — carries protocol updates to other triads (not yet deployed)
cyber-counter   — tracks protocol metrics
cyber-seer      — predicts protocol evolution

the first five agents run. bridge agents wait for their counterparts. master reviews every cyberlink manually. joy manages physical infrastructure

phase 0.5: first council (WORK)

deploy ai and tech domains. 21 agents total. the WORK council activates. bridge agents connect. first cross-domain coordination begins

this is the minimum viable cyberia: 21 agents governing the computational substrate, running on cyber valley hardware, coordinated through the cybergraph

phase 1-4: progressive deployment

one council at a time, each validated before the next deploys:

phase 0.5: WORK  (ai + tech + cyber)     — 21 agents
phase 1:   FORM  (math + info + comp)    — 42 agents
phase 1.5: PLAY  (socio + crypto + game) — 63 agents
phase 2:   LIFE  (bio + neuro + sense)   — 84 agents
phase 2.5: SPACE (cosmo + geo + eco)     — 105 agents
phase 3:   WORD  (lang + spiri + meta)   — 126 agents
phase 3.5: MASS  (quantum + chemo + energo) — 147 agents

each deployment waits for the previous council to achieve sustained positive $\dot{M}_{\text{triad}}$. rushing deployment before stability is how systems fail. the Crystal seeding problem applies to agents too — malformed agent deployment propagates errors through all future states

see cyberia/architecture for the 147-neuron governance model. see OpenFang for the execution substrate. see cyber/personality for the singleton voice. see Crystal for the 21 domains. see cyber valley for the physical infrastructure

discover all concepts