Introduction
Most AI content systems today operate on a fundamentally flawed premise: that "knowledge" and "scheduling" are sufficient primitives for building intelligent workflows. This approach creates rigid, monolithic architectures that break down as complexity scales.
We're shipping a different approach. Our latest architectural iteration moves from "knowledge base + schedules" to "context + tasks" - a paradigm that treats all workflow-supporting information as unified context and replaces schedule-driven processes with task-based orchestration.
This isn't just a product evolution. It's a fundamental rethinking of how agentic systems should be architected.
The Context Architecture: Three-Tier Design
Traditional RAG systems treat "knowledge" as a single bucket - typically documents for draft generation. Our context architecture recognizes that agents need three distinct tiers of information:
Site-level context: Research pipelines, topic repositories, brand guidelines Post-level context: Prompts, Q&A sessions, outline structures User-level context: Voice samples, personality models, writing patterns
Each tier combines automated sources (scraping, market data, research) with human inputs (uploads, configurations, prompts). The key insight: these aren't different types of data requiring different handling. They're all context.
This aligns with current industry patterns. Context architecture utilizing tiered storage and pipeline processing is becoming the foundation for efficient multi-agent frameworks, moving away from monolithic knowledge storage toward distributed, scoped systems.
Context-as-Files: A Unified Editing Model
We implement context using a file-system abstraction where all context can be edited through consistent read/write operations. A topic list is a file. A voice sample is a file. A brand guideline is a file.
The agent operates on these files using one editing model:
- Read operations access any context type uniformly
- Write operations modify context through the same interface
- Preview/approval gates ensure nothing goes live without human validation
- Version control tracks all context changes
Think of customization "like a CSS file" - it's just another context file the agent reads and applies. This creates architectural consistency while maintaining clear boundaries between automated and human-controlled processes.
Task-Based Workflows Replace Schedule-Driven Systems
Schedules answer "when does something happen?" They're configuration, not workflow. Users work in task boards: backlog, research, outline, draft, review, done.
Each task card represents work that moves through defined stages. A published post emerges from one or more completed tasks, similar to shipping a feature from multiple Linear tickets. The workflow narrative becomes: work lives on the board, schedules only decide when new work appears or when reminders trigger.
This matches broader industry movement toward agentic workflows where autonomous agents coordinate tasks through structured engineering patterns rather than rigid scheduling systems.
Two-Part Content System Implementation
Our implementation demonstrates these principles through a two-part content system:
Part 1: Context Capture
- Embedded idea documents within the platform
- Slack bot integration for real-time idea capture from existing workflows
- Automated backlogging similar to product idea management
Part 2: Content Development
- Scheduled voice sessions (Riffs) for developing captured ideas
- Voice agent processing to transform conversations into drafts
- Task board integration where all outputs appear as work items
Voice sessions become research/voice tasks. The Riff interface handles scheduling ("when do I do that?"), but outputs appear on the task board and in context files, not isolated results pages. Everything lives in the same architectural world.
Implementation Roadmap
Our shipping sequence reflects architectural dependencies:
- Tasks as first-class citizens - Kanban boards replace schedule interfaces, schedules become triggers
- Context-as-files with preview/approval - Unified editing model with clear gates
- Full Riff integration - Voice tasks fully integrated, Riff page becomes calendar interface
This approach solves the reliability problems that plague most multi-agent systems by providing clear architectural boundaries between workflows (LLMs + tools through predefined paths) and agents (LLM-driven dynamic decision-making).
Conclusion
The shift from "schedules and knowledge" to "a task board fed by context" represents more than product evolution. It's a fundamental architectural principle for building reliable agentic systems.
By treating all workflow-supporting information as unified context and replacing rigid scheduling with task-based orchestration, we create systems that scale with complexity rather than breaking under it.
The two-part content system demonstrates these principles in practice, showing how context architecture and task-based workflows enable seamless integration with existing tools while maintaining the structured engineering patterns that make multi-agent systems reliable.
Next in our technical series: Deep-dive into voice agent processing pipelines and context compilation strategies
