title: ‘Model Taxonomy: Reasoning-Class vs Agentic-Class Selection Heuristic’ date: 2026-02-15 tags:

model
workflow options: token_size: 1007 id: 26027 status: accepted superseded_by: null type: adr version: 1.0.0 birth: ‘2026-02-15’ authors:
name: Vadim Rudakov email: rudakow.wadim@gmail.com description: Tool-Agnostic Model Selection Heuristic Based on the Reasoning-Class vs Agentic-Class Taxonomy..

ADR-26027: Model Taxonomy: Reasoning-Class vs Agentic-Class Selection Heuristic¶

Title¶

Tool-Agnostic Model Selection Heuristic Based on the Reasoning-Class vs Agentic-Class Taxonomy.

Status¶

Accepted

Supersedes: ADR-26006, ADR-26008

Date¶

2026-02-15

Context¶

ADR-26006 and ADR-26008 introduced a valuable model taxonomy — distinguishing Reasoning-Class models (optimized for abstract synthesis and deep logic) from Agentic-Class models (optimized for instruction adherence and tool-use precision). However, both ADRs bind this taxonomy to a specific toolchain (aidx/aider phases), creating an unnecessary coupling between a general selection principle and a particular orchestration tool.

The AI engineering landscape has evolved:

Convergent model capabilities: Modern frontier models increasingly combine strong reasoning with high instruction adherence, blurring the rigid two-class separation.
Tool-agnostic methodology: This repository teaches principles that apply regardless of whether the practitioner uses aider, Claude Code, Cursor, or a custom agentic system.
Selection heuristic, not phase mandate: The taxonomy is most useful as a heuristic for choosing the right model for a given cognitive task, not as a rigid mapping of model classes to pipeline phases.

Decision¶

We adopt the Reasoning-Class vs Agentic-Class taxonomy as a tool-agnostic selection heuristic for AI engineering workflows.

Two cognitive axes (not rigid classes):
- Reasoning axis: Abstract synthesis capacity, GPQA/AIME ceiling, ability to challenge assumptions and surface hidden architectural contradictions.
- Agentic axis: Instruction adherence, tool-use precision, deterministic output structure, ability to follow templates without drift.
Selection by task, not by pipeline phase: Choose models based on the cognitive demands of the task at hand:
- Discovery and requirements engineering → prioritize the reasoning axis.
- Structured plan generation and code execution → prioritize the agentic axis.
- Verification and audit → prioritize both axes (System 2 reasoning with structured output).
No tool coupling: This heuristic applies equally to web chat interfaces, CLI agents, IDE copilots, and custom orchestration systems. No specific tool is mandated.
Hybrid models acknowledged: Modern models may score high on both axes. The taxonomy remains useful as a lens for evaluating fitness, not as a binary classifier.

Consequences¶

Positive¶

Portable knowledge: Practitioners can apply this selection heuristic regardless of their toolchain.
Future-proof: As models evolve, the two-axis framework accommodates hybrid capabilities without requiring a new ADR.
Preserved value: The core insight from ADR-26006 and ADR-26008 — that different cognitive tasks demand different model strengths — is retained and generalized.

Negative¶

Less prescriptive: Practitioners must evaluate models against both axes rather than following a simple “use model X for phase Y” rule. Mitigation: The Model Classification notebook provides concrete examples and a decision matrix.

Alternatives¶

Retain ADR-26006 and ADR-26008: Rejected. Their coupling to aidx phases contradicts the repository’s tool-agnostic goals.
Abandon the taxonomy entirely: Rejected. The reasoning/agentic distinction remains a valuable selection heuristic even as models converge.

References¶

ADR-26006: Requirement for Agentic-Class Models for the Architect Phase (superseded)
ADR-26008: Selection of Reasoning-Class Models for Abstract Synthesis (superseded)
General Purpose vs Agentic Models — detailed classification and decision matrix

Participants¶

Vadim Rudakov
Claude Opus 4.6