Narrative Work OS
For teams
10–200 people who need shared memory, decision traceability, and async coordination.
Not a SaaS
A set of structured conventions and file formats. You own your data. It runs anywhere git runs.
AI-native
Designed from the start for AI agents to read, write, and operate within the same files as humans.
File Layer
All data is markdown files with YAML frontmatter. No proprietary formats.
Version Layer
Git provides full history, branching, and collaboration via PRs.
Agent Layer
AI agents read and write files like any contributor.
Human Layer
Oracles (humans) approve decisions, set direction, and audit agent work.
Narrative Layer
Optional. Guilds, lore, and game mechanics that make the system engaging for humans.
L4 (Narrative) is opt-in. The system works without it. It adds engagement and cultural coherence for teams that want it.
Mission System
Each unit of work is a structured document with a unique ID, acceptance criteria, epistemic value, and pragmatic value. Missions are versioned in git.
Why it matters
Transforms tasks into documented knowledge. Closing a mission leaves a trace — what was done, why it diverged from the plan, and what was learned.
vs. alternatives
Linear and Jira track completion. The Mission System tracks knowledge.
Fields: id · title · type (biological/digital/hybrid) · priority · effort · status · story · acceptance criteria · epistemic value · pragmatic value · execution reality
Blueprints (System Maps)
Living architecture documents that show the current state, target state, gap delta, and open questions for each subsystem.
Why it matters
Most organizations have architectural decisions in people's heads. Blueprints externalize that knowledge into auditable, updatable files.
vs. alternatives
Architecture Decision Records (ADRs) capture one decision. Blueprints show the whole system at a glance.
Fields: id · area · status semaphore (green/yellow/red) · current state · target state · related decisions · gap → mission delta table · open questions · dependencies
Decision Registry
Append-only log of architectural and strategic decisions. Each record captures context, the decision made, alternatives rejected, and pros/cons.
Why it matters
Decisions made without documentation get remade. The registry prevents wheel reinvention and makes the cost of changing direction explicit.
vs. alternatives
Notion docs get edited and lose history. The Decision Registry is immutable — superseded by new decisions, never deleted.
Fields: id · title · date · status (active/provisional/superseded) · context · decision · why · rejected alternatives · pros/cons
Digital Agents (CAO)
AI agents with persistent identity files (SOUL.md, OPERATOR.md, MEMORY.md) that operate within the system. Each agent has a guild, a role, and operational laws.
Why it matters
Agents are not chatbots. They are long-running collaborators with memory, responsibilities, and verifiable identity. They read and write the same files humans do.
vs. alternatives
ChatGPT and Copilot are stateless. NWOS agents maintain state across sessions via the git repository.
Fields: SOUL.md (identity) · OPERATOR.md (laws) · MEMORY.md (persistent context) · guild assignment · mission assignments · operation logs
Operational Reports
Daily and weekly markdown reports committed to the repository. Reports are append-only — closed reports are never modified.
Why it matters
Reports create accountability without overhead. A 5-minute daily summary in git is auditable, searchable, and doesn't require a BI tool.
vs. alternatives
Slack summaries disappear. Email reports aren't searchable. Git reports are permanent and diffable.
Fields: id · date · agent · model · completed work · epistemic value · pragmatic value · pending items requiring human input
Protocols
Operational procedures written as markdown files. Protocols define how recurring tasks are executed — from briefing an agent to onboarding a new member.
Why it matters
Recurring tasks without protocols become dependent on specific people. Protocols make processes portable.
vs. alternatives
SOPs in Confluence rot and go unread. Markdown protocols are lightweight, versioned, and executable by agents.
Fields: id · title · trigger · steps · owner · version
File over App
Data lives in markdown files, not locked in a SaaS database. If the tool disappears, the knowledge survives.
Append-only history
Decisions and reports are never edited retroactively. The record of what was thought and done at each moment is permanent.
Explicit over implicit
Every architectural decision, every divergence from a plan, every open question is written down. Nothing lives only in someone's head.
Epistemic value as first-class citizen
Every mission asks: what do we learn by doing this? Documentation that only records what was done is incomplete.
Human approval for external actions
Agents execute internal work autonomously. Any action that leaves the system (email, publish, deploy) requires explicit human approval.
Layered architecture
Each layer (files, version control, agents, humans, narrative) can fail without taking the others down. The system degrades gracefully.
Requirements
Git repository
GitHub, GitLab, or self-hosted
Markdown editor
Any editor that reads .md files
AI agent access
Optional. Unlocks the agent layer.
Reference implementation
The repository numengames/numinia-digital-agents is the live reference implementation of the NWOS, licensed CC0.
About the Narrative Layer
The NWOS was built alongside Numinia — a narrative universe where guilds, missions, and archives are not metaphors but the actual vocabulary of work. The narrative layer is not required, but it changes how people relate to the system.
If you want the system without the narrative, it's still the same files, the same git workflow, the same agents. The narrative is an optional skin that makes the system feel alive.