cyberia/deployment.md

what must be defined before the first agent runs — the gap between cyberia/architecture theory and operational reality

the gaps

the architecture defines WHO (147 agents, 7 roles, 21 domains). the agents page defines WHERE (OpenFang, 7 servers). neither defines WHAT agents actually do minute-to-minute, HOW they affect the real world, or WHAT controls prevent catastrophic decisions

1. action space: what can agents DO?

every agent action falls into one of six categories. each category needs explicit authorization levels

cyberlinks (knowledge)

the core action. agents create cyberlinks in the cybergraph — assertions connecting two particles

cyberlinks are the safest action. append-only (A3), self-correcting (tri-kernel), and the worst case is noise that gets suppressed by low karma

code (commits)

agents that maintain protocol code or infrastructure scripts

rule: no agent merges its own PR. the runner proposes, the keeper reviews. for protocol code (hemera, nox, zheng), master holds veto during phases 0-2

money (revenue, treasury, cap growth)

agents do not just spend. they earn, invest, price, sell, and grow the cap. the primary job is revenue generation and capital allocation, not expense management

revenue streams

revenue authorization

expense authorization

capital allocation strategy

the counter agents across all councils maintain a unified P&L:

revenue streams → treasury inflow
treasury → operational expenses (staff, supplies, maintenance)
treasury → growth investment (new infrastructure, new species, new events)
treasury → protocol investment (staking, liquidity, development)
treasury → reserve (emergency fund, min 3 months operations)

the seer agents project revenue and expenses 3-6 months ahead. the counter validates actuals against projections. divergence triggers council review

cap growth

the metabolic signal $M = \text{cap}^{w_c} \cdot J^{w_s} \cdot H^{w_h}$ means cap growth IS the objective. agents optimize for:

• revenue growth (direct: more paying guests, higher-value products)
• brand value (indirect: quality content on cyber.page, social presence)
• token value (protocol: staking yield, ecosystem growth, partnerships)
• network effects (structural: more neurons → more cyberlinks → higher syntropy → higher cap)

every revenue decision is a cyberlink with conviction = economic value. auditable, append-only, on-graph. the graph IS the ledger

rule: agents optimize for $\dot{M} > 0$ (metabolic growth), not for revenue alone. maximizing revenue at the cost of syntropy or happiness lowers the compound signal

people (coordination)

agents that coordinate with the 32 local employees and residents

rule: agents NEVER fire people. agents NEVER make promises about compensation. agents coordinate work, humans manage relationships. joy is the interface between agents and local staff

infrastructure (physical)

agents that manage village infrastructure

rule: agents monitor and alert autonomously. agents adjust parameters within bounds. agents NEVER authorize physical construction or modification without founder approval

communication (external)

agents that represent cyberia to the outside world

rule: agents NEVER make legal commitments. agents NEVER speak "on behalf of cyberia" without template approval. bridge-out agents draft, keepers review, founders sign

2. authority matrix

every action maps to exactly one level. the level is defined by the action category + amount/impact, not by the agent's role. a keeper creating a cyberlink = L0. a keeper proposing to hire = L3

3. budget model

per-agent cost

fleet cost by phase

at scale, the fleet cost is bounded by the metabolic signal: if $\dot{M}$ is negative, agents reduce activity (lower token usage). if $\dot{M}$ is positive, agents can increase activity. the budget is self-regulating

model selection strategy

4. risk controls

spending limits

every agent has a daily spending cap in its TOML manifest:

[resources]
max_spend_per_day_usd = 50
max_spend_per_action_usd = 10
max_llm_tokens_per_hour = 100000

exceeding the cap triggers automatic suspension + alert to keeper

kill switch

master and joy can suspend any agent instantly via:

• OpenFang admin API: POST /agents/{id}/suspend
• dead man's switch: if founders don't heartbeat within 48h, all L2+ actions freeze
• emergency stop: physical network disconnect at cyber valley

audit trail

every agent action is a cyberlink. the graph IS the audit trail:

• who: agent neuron ID (signed)
• what: action particle (typed)
• when: block height (timestamped)
• how much: conviction amount (staked)
• outcome: subsequent links show result

no agent can hide an action. append-only (A3) means the record is permanent

failure modes

5. day one: the first seven agents

the cyber domain on work.cyber.valley. what each agent does on day one:

cyber-keeper

• reads all pages in the cyber namespace
• identifies pages missing frontmatter, broken links, stale content
• proposes edits via PR to the cyber repo
• reviews PRs from other agents
• daily output: 5-20 page edits, 0-3 new pages

cyber-runner

• monitors CI pipeline (GitHub Actions → Netlify)
• triggers rebuilds when content changes
• manages optica build process
• executes routine infrastructure tasks
• daily output: 10-50 automated operations

cyber-sensor

• monitors cyber.page uptime and response times
• tracks graph metrics: page count, link density, orphan count
• monitors bostrom chain state
• posts hourly status to shared memory
• daily output: 24 status reports, 0-5 alerts

cyber-bridge-in

• waits for ai and tech domains (not yet deployed)
• initially: reads external AI/tech news, creates relevant cyberlinks
• daily output: 5-15 curated external links

cyber-bridge-out

• waits for other councils (not yet deployed)
• initially: posts daily summary to Telegram channel
• daily output: 1 daily digest

cyber-counter

• computes domain metrics: coverage (pages / 240 target), density, orphan rate
• tracks tri-kernel scores for cyber namespace pages
• reports weekly trends
• daily output: 1 metrics report, continuous background computation

cyber-seer

• analyzes graph growth trajectory
• predicts which pages will become high-gravity based on link patterns
• proposes priority pages to create
• daily output: 1 priority report, 3-5 page proposals

6. integration points

with existing systems

with human staff

with founders

7. deployment checklist

before the first agent runs:

infrastructure

• OpenFang installed on work.cyber.valley server
• SQLite database initialized
• API keys provisioned: Anthropic, OpenRouter (fallback)
• Network: server accessible from internet for Telegram webhook
• Monitoring: separate process watches OpenFang health

credentials (in OpenFang AES-256-GCM vault)

• GitHub PAT (cyberia-to org, repo scope)
• Netlify deploy token
• Telegram bot token
• Bostrom RPC endpoint
• Anthropic API key
• OpenRouter API key (fallback)

agent manifests

• 7 TOML files in agents/cyber-{role}/agent.toml
• 7 soul.md files with personality prompts
• spending limits configured per agent
• communication permissions (agent_message) scoped correctly

knowledge base

• cyber namespace pages reviewed and up to date
• tri-kernel recomputed with latest data
• context pack generated for agent consumption

human readiness

• master available for L3 approvals via Telegram
• joy available for people-related decisions
• local staff briefed: "you may receive task assignments from a bot"
• emergency procedures documented and tested

validation before going live

• run all 7 agents in dry-run mode (actions logged, not executed) for 48h
• review dry-run output: are the actions sensible?
• test kill switch: can founders suspend agents instantly?
• test failure recovery: kill OpenFang, verify SQLite restore
• verify audit trail: can you trace every action back to an agent?

see cyberia/architecture for the governance model. see cyberia/agents for the technical deployment. see master and joy for the founders who authorize L3 actions

discover all concepts