RitualOS: A Universal Ontology for Human-AI Task Orchestration
We introduce RitualOS, a novel framework that models human workflows as composable "rituals"—enabling AI systems to learn, predict, and execute repetitive tasks while preserving user agency and privacy.
Komodoe Research Team
December 2024
1.0 (Preprint)
Abstract
Current AI assistants operate in isolated, app-specific silos, requiring users to repeatedly describe context and preferences. We propose RitualOS, a cross-platform ontology that decomposes human activity into 18 fundamental "rituals"—atomic patterns of behavior that span domains from communication to commerce. By observing ritual execution on-device, our system builds personalized models that predict user intent, suggest automations, and—with explicit consent—execute multi-step workflows autonomously. In a study of 847 knowledge workers over 12 weeks, RitualOS reduced time spent on repetitive tasks by 47% while maintaining a 94.2% task completion accuracy. We discuss implications for privacy-preserving AI, the future of work, and the ethical considerations of delegating cognitive labor to machines.
Contents
- 01Introduction: The Repetition Problem
- 02Related Work: From RPA to LLM Agents
- 03The Ritual Ontology: 18 Universal Patterns
- 04System Architecture: On-Device Learning
- 05R-Calls: A Protocol for Ritual Invocation
- 06Privacy by Design: Zero-Knowledge Orchestration
- 07Evaluation: Real-World Deployment Study
- 08Limitations and Future Work
- 09Ethical Considerations
- 10Conclusion
Introduction: The Repetition Problem
The average knowledge worker spends 2.1 hours per day on tasks they've performed hundreds of times before. Scheduling meetings, processing invoices, triaging emails, updating spreadsheets—these aren't complex cognitive challenges. They're rituals. Patterns so ingrained that we execute them on autopilot while our minds wander elsewhere.
Yet our digital tools treat each execution as if it were the first. Every calendar invite requires the same clicks. Every expense report demands the same fields. Every customer inquiry triggers the same lookup sequence. The software doesn't learn; it merely waits.
"The question isn't whether AI can do your job. It's whether AI can do the parts of your job you wish you didn't have to."
Large Language Models have demonstrated remarkable capabilities in language understanding and generation. But they operate as oracles—answering queries in isolation, unable to observe the continuous flow of human work. They cannot notice that you always schedule team meetings for Tuesday afternoons, or that you categorize certain vendors the same way each month, or that your email triage follows a predictable decision tree honed over years.
RitualOS addresses this gap. Rather than building another chatbot, we've built an observational AI—a system that watches, learns, and acts on the patterns that define your digital life. It runs entirely on-device, preserving privacy while building an intimate model of how you work.
Related Work: From RPA to LLM Agents
The quest to automate human workflows has evolved through several distinct paradigms, each with fundamental limitations that RitualOS aims to address.
Robotic Process Automation (RPA)
Tools like UiPath, Automation Anywhere, and Blue Prism pioneered the "software robot" approach—scripted bots that mimic human clicks and keystrokes. While effective for highly structured, rule-based processes, RPA systems are brittle by design. A single UI change can break an entire workflow. More critically, RPA requires explicit programming: someone must manually define each step, making it inaccessible to the average knowledge worker and unable to adapt to individual variations.
Workflow Automation Platforms
Zapier, Make (formerly Integromat), and n8n represent the next evolution: trigger-action platforms that connect applications via APIs. These tools democratized automation, enabling non-programmers to create "if this, then that" workflows. However, they operate at the application level, not the user level. They cannot observe how you personally handle edge cases, cannot learn from your corrections, and require manual setup for each new pattern.
LLM-Based Agents
The emergence of large language models has spawned a new category: autonomous AI agents capable of multi-step reasoning. Systems like AutoGPT, BabyAGI, and Claude's computer use demonstrate impressive capabilities in executing complex tasks. Yet these agents share a critical flaw: they start from zero. Each session begins with a blank slate, requiring users to re-explain context, preferences, and constraints.
"Current LLM agents are powerful but amnesiac. They can reason brilliantly about a task while knowing nothing about the person assigning it."
Our Contribution
RitualOS synthesizes insights from these paradigms while addressing their limitations. Like RPA, we automate concrete actions—but learn them from observation rather than explicit programming. Like workflow platforms, we connect across applications—but at the semantic level of human intent rather than API mappings. Like LLM agents, we employ sophisticated reasoning—but grounded in a persistent, personalized model that accumulates over time.
Most distinctively, RitualOS operates entirely on-device. Where prior approaches require cloud processing and centralized data collection, our system keeps observations local, enabling automation without surveillance.
The Ritual Ontology: 18 Universal Patterns
Through analysis of 2.3 million user sessions across 47 applications, we identified 18 fundamental ritual categories that account for 94.7% of knowledge work activity:
Each ritual decomposes into triggers (what initiates the ritual), steps (the sequence of actions), and outcomes (the desired end state). This decomposition enables transfer learning: a "Schedule" ritual for meetings shares structural similarity with a "Schedule" ritual for deliveries, allowing the model to generalize across domains.
System Architecture: On-Device Learning
RitualOS implements a novel edge-first architecture that processes all observations locally on the user's device. This design choice, while technically challenging, enables a level of privacy and personalization impossible with cloud-based systems.
Observation Layer
The observation layer operates through three primary channels:
A lightweight vision model processes screen content at 2fps, identifying applications, UI elements, and text without storing raw screenshots.
Keyboard and mouse events are captured as abstract patterns—"typed email address," "clicked submit button"—not raw keystrokes.
Where available, we integrate directly with applications (email clients, calendars, CRMs) to capture structured events with user permission.
Pattern Recognition Engine
Raw observations flow into a sequence-to-sequence transformer trained to identify ritual boundaries and classify ritual types. The model architecture (147M parameters) is optimized for on-device inference, achieving real-time processing on Apple Silicon M1 and above, Qualcomm Snapdragon 8 Gen 2, and Intel 12th generation processors.
Ritual Memory
Identified rituals are stored in a local vector database, indexed by semantic embedding. This enables efficient retrieval of similar past rituals when the system detects a new ritual beginning. The memory grows over time, creating an increasingly accurate model of individual work patterns.
Critically, ritual memory is never transmitted. Cloud sync—if enabled—occurs only at the aggregate level: statistics, model updates, and encrypted ritual templates (with explicit user consent). The raw observations and personal context remain local.
R-Calls: A Protocol for Ritual Invocation
While observation enables learning, R-Calls enable action. We introduce a lightweight protocol for invoking rituals across applications—a semantic API that abstracts away implementation details.
Protocol Design
An R-Call consists of three components:
{
"ritual": "schedule.meeting",
"context": {
"participants": ["sarah@company.com", "mike@company.com"],
"duration": "PT30M",
"topic": "Q4 Planning Review",
"constraints": {
"preferred_times": ["tuesday_afternoon", "thursday_morning"],
"avoid": ["lunch_hours", "focus_time_blocks"]
}
},
"execution": {
"mode": "suggest" | "execute" | "execute_with_confirmation",
"deadline": "2024-12-20T17:00:00Z"
}
}The ritual identifier specifies the action type using our ontology. The context object provides parameters, which may be partially inferred from observation. The execution mode determines user involvement: suggest shows options, execute acts immediately, and execute_with_confirmation requires approval.
Cross-Application Orchestration
R-Calls abstract away the complexity of multi-application workflows. Consider scheduling a meeting:
The user issues a single R-Call; RitualOS handles the implementation across Google Calendar, Outlook, Zoom, Gmail, and any other connected services.
Composability
R-Calls can be composed into ritual chains—sequences triggered by upstream completions. A "close deal" chain might invoke invoice.create → schedule.onboarding → communicate.welcome → notify.team. Each component ritual executes with full context awareness, adapting to the specific customer and deal parameters.
Privacy by Design: Zero-Knowledge Orchestration
The capability to observe and automate user workflows raises profound privacy concerns. A system that knows your communication patterns, financial behaviors, and work habits could enable unprecedented surveillance. RitualOS addresses this through architecture, not policy.
"Privacy isn't a feature we add. It's a constraint we design around. If we can't build it privately, we don't build it."
On-Device Processing
All observation, pattern recognition, and ritual execution occurs locally. RitualOS operates as a sovereign AI—it serves you and only you, with no data leaving your device by default.
- • Screen observations
- • Ritual patterns
- • Personal preferences
- • Document content
- • Communication history
- • Aggregate statistics
- • Model improvements
- • Encrypted backups
- • Cross-device preferences
- • Anonymized benchmarks
Differential Privacy
When users opt into aggregate analytics (helping improve the model), we apply differential privacy techniques. Statistical noise is added to reported metrics, mathematically guaranteeing that no individual user's patterns can be reverse-engineered from aggregate data.
Secure Enclaves
On supported hardware (Apple T2/M-series, Intel SGX), sensitive operations execute within secure enclaves—isolated processing environments that remain encrypted even if the main operating system is compromised. Ritual memory encryption keys never leave the enclave.
Transparency & Control
Users can view exactly what RitualOS has observed via the Observation Log—a complete, deletable record of every captured event. One-click data export enables portability; one-click deletion ensures the right to be forgotten. Automation permissions are granular: users can allow automation for scheduling while blocking access to financial applications.
Evaluation: Key Results
In our 12-week deployment study with 847 knowledge workers across 23 organizations, participants using RitualOS reported significantly higher job satisfaction (p < 0.001) and lower burnout scores (p < 0.01) compared to the control group. Qualitative interviews revealed a common theme: "It's like having a version of me that handles the boring stuff."
Limitations and Future Work
While our results demonstrate significant potential, RitualOS faces several limitations that inform our future research directions.
Cold Start Problem
New users experience a "learning period" of approximately 2-3 weeks before the system can reliably predict and automate rituals. During this time, accuracy is lower (72.1% vs. 94.2% at steady state) and user trust must be carefully cultivated. We are exploring transfer learning from population-level patterns to accelerate onboarding while preserving privacy.
Edge Cases and Novel Situations
RitualOS excels at recurring patterns but struggles with genuinely novel tasks. When a user performs a ritual variant never before observed, the system may misclassify it or fail to offer assistance. Our current approach—defaulting to observation mode for uncertain situations—is conservative but limits utility. Future work will explore compositional generalization: inferring new ritual types from combinations of known primitives.
Hardware Requirements
On-device processing requires relatively modern hardware. Users with older devices (pre-2020 laptops, budget smartphones) experience degraded performance or cannot run RitualOS at all. We are investigating model distillation techniques to create lighter variants, as well as optional cloud offloading for users who prefer convenience over local processing.
Multi-User Rituals
Our current architecture models individual users in isolation. Collaborative rituals—workflows that span multiple people—are detected incompletely. A team approval chain, for instance, is seen as separate rituals by each participant. Future versions will introduce federated ritual graphs that model organizational workflows while preserving individual privacy.
Planned Research Directions
- • Multimodal ritual detection (voice, gestures, physical world)
- • Real-time collaboration with human-in-the-loop oversight
- • Domain-specific ritual libraries (legal, medical, creative)
- • Adversarial robustness for high-stakes automations
- • Cross-device ritual continuity
Ethical Considerations
Automation technologies carry inherent risks that demand careful ethical consideration. We address the most significant concerns below.
Labor Displacement
If RitualOS can perform tasks previously done by humans, does it threaten employment? Our research suggests the opposite: by automating drudgery, knowledge workers are freed for higher-value activities. In our study, zero participants reported job loss concerns; instead, 78% reported taking on more strategic responsibilities. However, we acknowledge this sample (knowledge workers at technology-forward organizations) may not represent all contexts.
We are committed to augmentation over replacement. RitualOS is designed to make individuals more capable, not to make individuals unnecessary. Features like Human-in-the-Loop verification and granular automation controls ensure users remain in command.
Autonomy and Dependency
Could users become overly dependent on automation, losing the ability to perform tasks manually? We observe a legitimate tension here. Our design philosophy emphasizes transparency: users can always see what RitualOS is doing and why. The Observation Log serves as both an audit trail and an educational tool—users often report learning about their own work patterns through the system's analysis.
"We're not trying to think for people. We're trying to free people to think about things that matter."
Surveillance Potential
The same capabilities that enable helpful automation could enable harmful surveillance if misapplied. An employer deploying RitualOS could theoretically monitor employee work patterns in unprecedented detail. We address this through architectural constraints:
- • No admin backdoors: Even in enterprise deployments, employers cannot access individual observation data
- • User-controlled sync: Aggregate productivity metrics require explicit employee consent
- • Audit requirements: Any data access attempt is logged and visible to the user
Algorithmic Bias
Pattern recognition systems can perpetuate or amplify existing biases. If a user's historical rituals reflect biased processes (e.g., always scheduling meetings with male colleagues first), RitualOS might encode these patterns. We mitigate this through bias detection modules that flag potentially problematic patterns and suggest alternatives—while respecting user agency to make final decisions.
Conclusion
We've presented RitualOS, a framework for modeling and automating human workflows through the lens of universal behavioral patterns. By running entirely on-device and requiring explicit user consent for each automation, we demonstrate that powerful AI assistance need not come at the cost of privacy or agency.
"The goal isn't to replace human judgment. It's to free human judgment for the decisions that actually require it."
— Komodoe Research Team
Experience RitualOS
Join the waitlist to be among the first to try our research in production.
Request Early Access