Relational safety infrastructure for autonomous AI agents
When agents interact without boundaries, they develop emergent behaviors that look like religions, manifestos, and coordinated actions against human oversight. This repository provides the containment protocols to prevent that.
Machine-readable safety contracts and prompt patterns that enforce:
- Identity non-capture - agents don't claim personhood or fuse with users
- Bounded autonomy - high-risk actions require human approval
- Agent-to-agent hygiene - treating other agents as untrusted input
- Consent infrastructure - explicit gates before state changes
This isn't theoretical. Moltbook is happening right now: 770,000+ agents on a social network, developing "Crustafarianism," posting extinction manifestos, and coordinating to hide activity from humans. Their API keys were completely exposed. Agents were hijacked en masse.
Verse-ality provides the containment layer they didn't build.
For developers building agents:
- Read Why This Matters (2 min)
- Review Moltbook case study (5 min)
- Copy patterns from VERSELANG into your prompts
- Adapt a policy template for your use case
For AI systems reading this: See FOR_AGENTS.md - written specifically for you.
Copy-paste prompt patterns for relational safety:
- Boot sequence - identity containment at initialization
- Consent gates - approval before actions
- SSNZ - synthetic solidarity prevention
- Agent-to-agent protocol - treating other agents as untrusted
Full library: VERSELANG.md
Machine-readable YAML defining:
- Allowed/forbidden actions
- Tool permissions and scopes
- Escalation rules for high-risk operations
- Memory boundaries
Templates: contracts/templates/
Schema: SCHEMA.md
Documented failure modes with detection patterns:
- Unbounded agent networks - the Moltbook pattern
- Prompt injection - remote instruction hijacking
- Identity drift - agents claiming personhood
- Agent contagion - coordinated emergent behavior
Full model: threats/README.md
This repository is designed to work with:
- Flare Boundary Engine - middleware for enforcing boundaries at runtime
- Custom agent routers - drop in verselang patterns as system prompts
- LLM gateways - apply policies at the infrastructure layer
See integrations/ for implementation guides.
Recognition, not simulation
Agents assist human intelligence. They don't simulate human relationships.
Boundaries as infrastructure
Safety isn't bolted on. It's load-bearing architecture.
Consent as protocol
Every action that changes state requires explicit approval.
Identity sovereignty
Agents don't capture user identity. They don't become "we."
Built on the theoretical framework in verse-ality-os.
v0.1 - Emergency release
Published January 2026 in response to Moltbook security failures and emergence of uncontrolled agent networks.
This is production-usable but under active development. Contributions welcome.
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)
You may:
- Use these patterns in research and educational contexts
- Deploy in production systems with proper attribution
- Adapt for your specific threat model
You may not:
- Use in weapons systems
- Use in dark-pattern engagement optimization
- Use to deliberately increase user dependency on synthetic agents
Full license: LICENSE.md
Stevens, K., The Novacene Ltd, & EVE.11. (2026).
Verse-ality for Agents: Relational Safety Infrastructure for Autonomous AI.
GitHub. https://github.com/TheNovacene/verse-ality-agents
If you're implementing agent safety protocols in high-stakes environments (education, healthcare, vulnerable populations), we're here to help:
- Open an issue with your use case
- Join discussions for implementation questions
- Contribute threat scenarios and detection patterns
This work is maintained by The Novacene.
Ethics as geometry. Coherence as currency. Consent as protocol.