Overview
Extend Binding entry conditions beyond key existence to include capability matching, confidence thresholds, worker load, and quality-of-match scoring for dynamic agent selection.
Part of epic #77 (Blackboard Architecture Evolution).
Background
Recent LLM blackboard research demonstrates that dynamic agent selection — choosing which agent activates based on current context and agent fit — significantly improves solution quality and reduces wasted computation.
Current Binding conditions: a JQ expression that evaluates to true/false against CaseContext. This is binary — condition matches or it doesn't. No notion of quality-of-match, confidence, or capability fit.
What to build
- Confidence-weighted activation: Workers can optionally return a
confidence score (0–1) given current CaseContext state. `WorkerSelectionStrategy` uses scores to prefer higher-confidence candidates.
- Capability-based criteria: Binding conditions can include capability tags (e.g. `capability:nlp`, `capability:code-review`) resolved against registered Worker capabilities.
- Load-aware selection: `WorkerSelectionStrategy` can deprioritise Workers currently executing on other cases (worker load signal, already partially supported via `LeastLoadedStrategy`).
- Fit scoring API: `Worker.score(CaseContext)` — optional method returning activation fitness. The Binding evaluator calls it when present; returns 1.0 by default (current behaviour).
Relationship to existing work
Extends the expression-based Binding evaluation and `WorkerSelectionStrategy` SPI (issue #76). Binding conditions remain expression-based; this adds a scoring layer on top of eligibility evaluation.
Note: `WorkerSelectionStrategy` in the `WorkBroker` subsystem addresses a related but separate concern (worker selection for request-response tasks, not CaseContext-driven Binding activation).
Overview
Extend Binding entry conditions beyond key existence to include capability matching, confidence thresholds, worker load, and quality-of-match scoring for dynamic agent selection.
Part of epic #77 (Blackboard Architecture Evolution).
Background
Recent LLM blackboard research demonstrates that dynamic agent selection — choosing which agent activates based on current context and agent fit — significantly improves solution quality and reduces wasted computation.
Current Binding conditions: a JQ expression that evaluates to true/false against CaseContext. This is binary — condition matches or it doesn't. No notion of quality-of-match, confidence, or capability fit.
What to build
confidence score (0–1)given current CaseContext state. `WorkerSelectionStrategy` uses scores to prefer higher-confidence candidates.Relationship to existing work
Extends the expression-based Binding evaluation and `WorkerSelectionStrategy` SPI (issue #76). Binding conditions remain expression-based; this adds a scoring layer on top of eligibility evaluation.
Note: `WorkerSelectionStrategy` in the `WorkBroker` subsystem addresses a related but separate concern (worker selection for request-response tasks, not CaseContext-driven Binding activation).