evolution-taxonomy-classifier

Evolution Taxonomy Classifier

Overview

Your agent has diagnosed a reasoning failure. The execution graph shows where it went wrong and the cognitive-fault isolator has classified the breakdown. Now what? The temptation is to jump straight to a fix: tweak a prompt, retrain an adapter, add a guardrail. But self-evolution spans more than a single lever: it is a four-dimensional design space, and pulling the wrong lever wastes compute, introduces regressions, or both.

This skill locates any proposed evolution on the four Gao et al. axes:

• WHAT evolves — model (weights or prompts; needs execution graphs for causal tracing) | context (retrieval; this IS graph evolution: rewire edges, merge nodes, prune subgraphs) | tool (rewires the tool subgraph by reweighting task-type-to-tool edges on observed success) | architecture (graph surgery on the workflow graph itself).
• WHEN it fires — intra_test_time (within one request; must be sub-second; Reflect-Retry-Reward) | inter_test_time (between requests; can afford fine-tuning or graph restructuring; SEAL overnight, semantic backpropagation).
• HOW the agent learns — reward_based (scalar signals: InfoGain, user satisfaction) | imitation_based (copy successful trajectories) | population_based (maintain variants, select fittest).
• WHERE it applies — general_purpose (all tasks) | domain_specialized (one vertical, e.g. cascade failures in microservice topologies).

Every axis requires graph structure to operate. As the chapter states: the graph is not optional infrastructure here, it is the substrate that makes any of these axes operable. Alshikh's production research reinforces the point: the first methodology is GNN-inspired, where "each adaptation becomes a traceable node." The classifier attaches that graph-dependency rationale to every value it assigns, and route_failure maps a diagnosed failure to its axis per Table 7-1.

When to Use

• AFTER a diagnostic report exists and you are deciding which evolution lever to pull
• To sanity-check a proposed evolution: is this really model evolution, or is it context evolution wearing a model-evolution costume?
• To route a diagnosed failure type (FORMAT, REASONING, KNOWLEDGE) to its primary axis, timing, and mechanism
• When designing a self-evolution loop and you need the four-axis vocabulary to keep intra-test-time and inter-test-time paths distinct