Root Signals

Provides MCP tools that allow AI agents to evaluate their own outputs against developer-defined scoring rubrics. Agents can invoke evaluation endpoints that apply multi-dimensional scoring criteria (accuracy, relevance, completeness, etc.) to generated content, receiving structured feedback scores and reasoning. This enables agents to assess quality before returning results to users or triggering refinement loops.

Collects structured signals about agent execution (success/failure outcomes, evaluation scores, latency, token usage, error types) and logs them to a centralized signal store. Agents can emit signals at key decision points, and the system aggregates these signals to build performance profiles. This creates a telemetry foundation for understanding agent behavior patterns and identifying improvement opportunities.

Enables agents to use evaluation signals and performance data to automatically refine their behavior across multiple iterations. Agents can inspect their own evaluation results, identify failure patterns, and adjust their approach (prompts, tool selection, parameter tuning) before retrying tasks. The system tracks refinement iterations and measures improvement, creating a self-improving agent loop without human intervention.

Supports evaluation rubrics with multiple independent scoring dimensions (e.g., code correctness, readability, performance, security) where each dimension has its own scoring scale and criteria. Rubrics are defined as structured schemas that specify dimension names, scoring ranges, and evaluation instructions. The evaluation engine applies all dimensions to a single output and returns a multi-dimensional score vector, enabling nuanced quality assessment beyond single-metric scoring.

Exposes Root Signals evaluation and refinement capabilities as standard MCP tools that agents can discover and invoke like any other tool. The MCP integration layer handles tool schema definition, parameter validation, and response formatting, allowing agents to call evaluation and signal emission functions using their native tool-calling mechanisms. This enables seamless integration into existing agentic frameworks without custom glue code.

Analyzes accumulated performance signals to identify patterns in agent behavior and automatically suggest or apply behavior adaptations. The system correlates evaluation scores, execution outcomes, and signal metadata to detect failure modes (e.g., 'agent fails on tasks with X characteristic'), then recommends behavior changes (prompt modifications, tool additions, parameter adjustments) to address identified patterns. Adaptations can be applied automatically or presented to developers for review.