Recoleta
GitHub
Source note

How Much Static Structure Do Code Agents Need? A Study of Deterministic Anchoring

Authors
Zhihao Lin, Mingyi Zhou, Yizhuo Yang, Li Li
Source
arXiv
Published
Collection
Software Intelligence
Code AgentsStatic AnalysisFault LocalizationRepository NavigationSwe Bench
Open arXivMarkdown source

Summary

CodeAnchor adds static program links to grep-first code agents as plain-text comments. The paper claims these comments make repository navigation more accurate, shorter, and more repeatable without changing the agent loop.

Problem

  • Grep-first code agents miss call graphs, inheritance, imports, configuration use, and data-flow links, so they often stop at nearby text matches instead of the function that needs work.
  • This matters because two runs on the same issue can visit different files and produce different outcomes, which makes failures hard to inspect and reproduce.
  • The paper studies how much static structure helps when the baseline is already strong: Codex reaches 83.2% Func@5 on SWE-bench Lite, compared with 59.5% for the graph-based LocAgent under the reported matched setup.

Approach

  • CodeAnchor runs lightweight static analysis offline, then inserts the discovered facts next to functions, classes, files, and configuration entries as normal comments.
  • The agent still uses ordinary text search and file reads. When it opens code, the nearby comments show links such as CALLS, CALLED_BY, IMPORTS, BASE, DERIVED, CONFIG_USAGE, DATA_DEP, IO_DEP, and TEST_REF.
  • The paper compares raw grep with Anchor-Topo, Anchor-Dense, and Anchor-Inv. Anchor-Topo adds call, inheritance, import, and containment links; Anchor-Dense adds config and data-flow hints; Anchor-Inv keeps inverse links such as CALLED_BY while dropping forward links.
  • The Python prototype uses PyCG for call graphs and AST passes for imports, containment, inheritance, config use, constants, I/O, and test links.

Results

  • On SWE-bench Lite, lightweight topology improves function localization by +2.2 percentage points on Func@5 and shortens navigation by 1.6 interaction rounds.
  • Tags increase structural link-following from 0.15-0.18 to 0.21-0.24 and roughly halve run-to-run variance in the reported stability study.
  • On medium-scale repositories, the paper reports a +3.4 percentage point Pass@1 gain for single-run reliability.
  • The added comments cost about 9.9% more input tokens on SWE-bench Lite, described as roughly 10% in the abstract.
  • Static graph construction times are 6.8s for pytest-7432 at 73k LOC, 21.1s for sklearn-15512 at 247k LOC, 58.4s for astropy-12907 at 341k LOC, and 133.4s for django-13658 at 367k LOC.
  • The paper reports that inverse-only links help hub-heavy repositories, while dense tags have diminishing returns unless the task depends on implicit configuration or data-flow links.

Link