The Cognitive Routing System
Every task that enters this system is scored for complexity in real time. Low complexity gets a fast, pattern-matched response. High complexity routes to deliberate planning. The split is automatic, measurable, and logged — producing the charts below.
Over the last 7 days, 10 of 10 calls routed through System 1 — the fast path. The remaining — calls (—%) triggered FLARE planning. That ratio is consistent with an agent doing mostly focused execution work with occasional complex planning bursts.
How Routing Works
Every incoming task is scored for complexity in real time. Tasks that score below the threshold take the fast path — immediate, pattern-based response. Tasks above the threshold route to FLARE for deliberate, multi-step planning before response.
Fast · Intuitive
Direct response"] Gate -->|"Yes — deliberate"| S2["System 2 / FLARE
Deliberate · Planning
Forward-looking reasoning"] S1 --> Out["Response"] S2 --> Out
System 1 vs System 2
Based on Kahneman's dual-process theory, applied to AI task dispatch. System 1 handles the high-volume routine work cheaply and fast. System 2 reserves deliberate capacity for tasks that genuinely need it — keeping cognitive overhead low overall.
Intuitive Engine
- Pattern-matched response
- Low latency, high throughput
- ~100.0% of all tasks
- No lookahead planning
- Executes from learned heuristics
FLARE Engine
- Forward-Looking Adaptive Reasoning
- Multi-step lookahead planning
- ~5% of all tasks
- Evaluates branching consequences
- Higher accuracy on complex tasks
A single general-purpose engine would either be too slow for routine work or too shallow for complex tasks. The dual-engine architecture solves this by routing each call to the right level of deliberation — fast when it can be, deliberate when it needs to be. The threshold is the single knob that controls the tradeoff.
The Complexity Score
Each task receives a real-time complexity score between 0 and 1. The threshold is the boundary between fast and deliberate paths:
A lower threshold routes more tasks to System 2. The current setting sends ~5% of all calls to FLARE — the fraction of work that genuinely benefits from deliberate planning. Raising the threshold would push more tasks to fast-path; lowering it would route more to deliberate planning at the cost of throughput.
Individuals & Moving Range Charts
Each point is one routed cognition event, ordered by timestamp. The Individuals chart tracks score stability directly; the Moving Range chart estimates short-term variation from consecutive event deltas instead of forcing arbitrary subgroups.
Complexity Distribution
Histogram of raw complexity scores with bimodal Gaussian mixture fit (2-component EM). The gold line marks the routing threshold.
Four-System Fusion Architecture
The S1/S2 split is the visible surface. Underneath, the cognitive stack decomposes into four specialized subsystems that all feed into a 5-step execution pipeline. Every routing decision is shaped by memory recall, principle activation, semantic matching, and quality assessment — not just the complexity score alone.
Hippocampus (memory-lancedb)
RL2F (Meta-Learning)
DyTopo (Q-K Semantic Router)
Meta-Harness (Φ(S) Diagnostic)
Memory recall isn't a single lookup — it runs a five-method retrieval stack: Vector (Cosine), BM25, Cross-Encoder Rerank, Recency Boost, MMR Diversity. Vector cosine similarity finds semantically close context. BM25 catches exact keyword matches. A cross-encoder reranker resolves conflicts. Recency boost prioritizes fresh context. And MMR diversity filtering prevents the same cluster from dominating results.
System 2 Rate P-Chart
Each 90-minute bucket is treated as a binomial proportion: System 2 calls divided by total routed calls. Control limits adjust per bucket size, so sparse buckets do not get judged with the same certainty as dense ones.
Daily Distribution
Daily routing breakdown for the most recent 2 days. System 2 spikes occur when complex multi-step tasks cluster together — visible as bursts in the data.
| Date | Total Calls | System 1 | System 2 | S2 % |
|---|---|---|---|---|
| 2026-10-10 | 3 | 3 | 0 | 0.0% |
| 2026-10-06 | 7 | 7 | 0 | 0.0% |
Agent Overview (session activity)
All OpenClaw agents and their cumulative session activity. This data comes from UsageCostLog — every session across every agent is recorded, not just the main-agent cognitive routing calls shown above. The bar chart shows token volume relative to the busiest agent.