Recoleta
GitHub
Source note

Inside the Scaffold: A Source-Code Taxonomy of Coding Agent Architectures

Authors
Benjamin Rombaut
Source
arXiv
Published
Collection
Software Intelligence
Coding AgentsSource Code TaxonomyAgent ArchitectureSoftware EngineeringLLM Scaffolds
Open arXivMarkdown source

Summary

This paper builds a source-code taxonomy of 13 open-source coding agent scaffolds. It argues that coding agents differ less by high-level labels like “planning” or “tool use” and more by concrete scaffold choices in control loops, tools, state, and context handling.

Problem

  • Research surveys group coding agents by abstract capabilities, but those labels cannot separate systems with very different scaffold code, cost, and failure modes.
  • Trajectory studies show what agents do at runtime, but they do not inspect the source code that drives those behaviors.
  • This matters because scaffold design affects reliability, token use, and evaluation, and current benchmark results often mix scaffold effects with model effects.

Approach

  • The authors analyze 13 open-source coding agents at pinned commit hashes, covering CLI tools, SWE-bench agents, and one minimal baseline.
  • They derive a taxonomy from source-code inspection rather than documentation, grounding each claim in file paths and line numbers.
  • Each agent is described across 12 dimensions in 3 layers: control architecture, tool and environment interface, and resource management.
  • The study identifies five reusable loop primitives: ReAct, generate-test-repair, plan-execute, multi-attempt retry, and tree search.
  • It compares how agents combine these primitives and how they differ in tool count, context compaction, state management, execution isolation, and model routing.

Results

  • The corpus includes 13 agents, selected from an initial pool of 22 candidates.
  • 11 of 13 agents combine multiple loop primitives instead of using a single control structure.
  • 7 of 13 agents use a sequential ReAct loop as their primary control structure.
  • Tool availability ranges from 0 tools in Aider to 37 action classes in Moatless Tools.
  • Context compaction spans 7 distinct strategies, and control designs range from fixed pipelines to full Monte Carlo Tree Search.
  • The paper does not report benchmark gains or new task-performance numbers. Its main concrete claim is that coding-agent scaffolds resist clean discrete categories and are better described as compositions along continuous design spectra.

Link