Building a Context-Aware Agent with Dual-Engine Architecture
AI Engineergo watch the original →
the gist
monday.com moves beyond simple RAG by using a dual-engine 'world model' that precomputes user context, separating long-term behavioral patterns from short-term urgent signals.
The Breakthrough
The authors shifted from a standard retrieval-augmented generation (RAG) approach to a precomputed 'world model' that separates user context into two distinct temporal engines, allowing the AI to understand the meaning and urgency of work rather than just indexing records.
What Actually Worked
- Implemented a dual-engine architecture: The system uses a 'slow engine' to mine weeks of activity for durable user profiles (work rhythms, goals, and collaboration patterns) and a 'fast engine' to process recent windows (last few days) for live signals like urgent tasks or overdue items.
- Modeled data as a graph of entities: Instead of raw text blobs, the system constructs a structured data model that maps relationships between Slack messages, meeting transcripts, and project board items to identify dependencies and blockers.
- Adopted a Lambda-style processing pattern: By merging precomputed batch data (the slow engine) with real-time stream processing (the fast engine), the system ensures the agent has a complete, served view of the user's state without needing to perform expensive, high-latency retrieval at query time.
- Enabled graceful degradation: Because the context is precomputed and verified against live data, the agent can fall back to the last known verified state if a specific data feed fails, preventing the agent from breaking entirely.
Context
Standard AI assistants often fail at prioritization because they treat every data point as equal, lacking the 'meaning' behind why a task exists or why a specific message matters. The authors argue that the bottleneck is not retrieval, but understanding. By treating the user's work history as a persistent data model—mirroring the hippocampus and neocortex in neuroscience—they allow their 'Sidekick' agent to reason about the user's actual priorities rather than just summarizing recent logs.
Notable Quotes
- "The problem was never the missing of data, the retrieval. The problem is the missing understanding."
- "A log never said what it mean."
- "The context is precomputed and served to their agent, so it degrades gracefully, judges when to speak up, and compounds as every new day and source sharpens the model."