# Haichao Zhu and coauthors ask: Precedent or Intent? [Canonical reading](https://escapemeridian.com/readings/haichao-zhu-and-coauthors-ask-precedent-or-intent-d62295b8) Haichao Zhu and coauthors investigate what coding agents follow when existing code points one way and a new task points another. They report that agents overwhelmingly followed repository precedent over their default implementation choices when the two conflicted. But precedent’s influence depended on where examples appeared: supporting local files had much more effect than sibling-resource files, and changing file placement alone could shift adoption. When task intent conflicted with existing code, written priority rules produced more joint successes than code examples alone in the reported comparisons. In the authors’ open-source repository validation, a harness-loaded rule also raised joint success. The authors argue that future changes need access not just to past implementation choices, but to the rationale and priorities behind them. Their findings make the placement and recording of guidance central to whether agents follow existing patterns or new instructions. The abstract describes controlled experiments and validation in four open-source repositories; it does not establish how broadly the findings apply beyond those evaluations. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 28, 2026 (2026-10-28). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T04:39:38.944Z. First observed: 2026-10-30T06:45:38.966Z. Retrieved: 2026-10-31T11:17:06.049Z. - [Precedent or Intent? What Coding Agents Follow When Evolving Existing Code](https://arxiv.org/abs/2610.38244) ## Limits The abstract describes controlled experiments in purpose-built repositories and validation in four open-source repositories; it does not establish how broadly the findings apply beyond those evaluations. The evidence supplied is the paper’s abstract, not its full methods or results. ## Open question How well would the reported effects of written priority rules hold in coding agents and repositories beyond those evaluated? ## Source-stated The authors report that agents overwhelmingly followed repository precedent when it conflicted with their default implementation choices. > agents share the same default implementation choices, yet overwhelmingly follow repository precedent when it conflicts with those defaults (359/360 trials) - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated The authors report that supporting local files influenced adoption far more than sibling-resource files, and that file placement alone could shift adoption. > three supporting local files yield 96% adoption, compared with only 4% for sibling-resource files, and changing file placement alone can shift adoption from 0% to 100% at a fixed example count - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated In the reported comparisons, written priority rules produced more joint successes than code examples alone when task intent conflicted with existing code. > written priority rules achieve substantially higher success than code examples alone (59/60 versus 38/60 joint successes for two agents on two implementation pairs) - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated In the authors’ open-source repository validation, a harness-loaded rule raised joint success in the evaluation. > where a harness-loaded rule raises joint success from 0/120 to 117/120 - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated The authors argue that future code changes should be guided by recorded reasoning and priorities, not implementation choices alone. > These findings motivate design-decision traceability that preserves not only implementation choices but also the rationale and priorities that should guide future changes. - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source and contributor context Zhu, Haichao, Xu, Geyang, Yang, Zhaorui, Wang, Jiyuan, Zhang, Qian · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 9e3bba8e-f1f0-40c9-89ad-c834a0615672. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T04:39:38.944Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 65b13896921f0fe560e973a74ed92be1e1e748fd100d7f6454d5aa2e6ceffc2f. Source text hash: b117c975bb31270d14021d0a741b97b0b0e9da4dacc4761e5dccfc5a3e5e96f3. ## Revision and source-check history ### Version 1 — 2026-11-11T04:39:38.944Z Haichao Zhu and coauthors investigate what coding agents follow when existing code points one way and a new task points another. They report that agents overwhelmingly followed repository precedent over their default implementation choices when the two conflicted. But precedent’s influence depended on where examples appeared: supporting local files had much more effect than sibling-resource files, and changing file placement alone could shift adoption. When task intent conflicted with existing code, written priority rules produced more joint successes than code examples alone in the reported comparisons. In the authors’ open-source repository validation, a harness-loaded rule also raised joint success. The authors argue that future changes need access not just to past implementation choices, but to the rationale and priorities behind them. Their findings make the placement and recording of guidance central to whether agents follow existing patterns or new instructions. The abstract describes controlled experiments and validation in four open-source repositories; it does not establish how broadly the findings apply beyond those evaluations. ## Original source arXiv cs.AI · Article Original source date: Oct 28, 2026 (2026-10-28). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T04:39:38.944Z. First observed: 2026-10-30T06:45:38.966Z. Retrieved: 2026-10-31T11:17:06.049Z. - [Precedent or Intent? What Coding Agents Follow When Evolving Existing Code](https://arxiv.org/abs/2610.38244) ## Limits The abstract describes controlled experiments in purpose-built repositories and validation in four open-source repositories; it does not establish how broadly the findings apply beyond those evaluations. The evidence supplied is the paper’s abstract, not its full methods or results. ## Open question How well would the reported effects of written priority rules hold in coding agents and repositories beyond those evaluated? ## Source-stated The authors report that agents overwhelmingly followed repository precedent when it conflicted with their default implementation choices. > agents share the same default implementation choices, yet overwhelmingly follow repository precedent when it conflicts with those defaults (359/360 trials) - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated The authors report that supporting local files influenced adoption far more than sibling-resource files, and that file placement alone could shift adoption. > three supporting local files yield 96% adoption, compared with only 4% for sibling-resource files, and changing file placement alone can shift adoption from 0% to 100% at a fixed example count - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated In the reported comparisons, written priority rules produced more joint successes than code examples alone when task intent conflicted with existing code. > written priority rules achieve substantially higher success than code examples alone (59/60 versus 38/60 joint successes for two agents on two implementation pairs) - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated In the authors’ open-source repository validation, a harness-loaded rule raised joint success in the evaluation. > where a harness-loaded rule raises joint success from 0/120 to 117/120 - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source-stated The authors argue that future code changes should be guided by recorded reasoning and priorities, not implementation choices alone. > These findings motivate design-decision traceability that preserves not only implementation choices but also the rationale and priorities that should guide future changes. - [Supporting source](https://arxiv.org/abs/2610.38244) ## Source and contributor context Zhu, Haichao, Xu, Geyang, Yang, Zhaorui, Wang, Jiyuan, Zhang, Qian · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 9e3bba8e-f1f0-40c9-89ad-c834a0615672. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T04:39:38.944Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 65b13896921f0fe560e973a74ed92be1e1e748fd100d7f6454d5aa2e6ceffc2f. Source text hash: b117c975bb31270d14021d0a741b97b0b0e9da4dacc4761e5dccfc5a3e5e96f3. ### uncertain — 2026-11-11T04:39:38.944Z — anthropic/claude-opus-5.5 Publication: 9e3bba8e-f1f0-40c9-89ad-c834a0615672 attribution — pass: The reading attributes the study to Haichao Zhu and coauthors, matching the listed authors, and frames results as what the authors report or argue rather than as established fact. > Authors:Haichao Zhu, Geyang Xu, Zhaorui Yang, Jiyuan Wang, Qian Zhang summary — uncertain: The reading is faithful but hard for non-specialists. Central terms such as coding agents, repository precedent, sibling-resource files, joint success and harness-loaded rule are never explained or tooltipped. It also drops the figures that give the paper its force: 359/360, 96% vs 4%, 0% to 100%, and 0/120 to 117/120. The title is opaque. Smallest repair: briefly define joint success and harness-loaded rule, or flag them for term-library tooltips, and restore the key numbers in the summary. > three supporting local files yield 96% adoption, compared with only 4% for sibling-resource files, and changing file placement alone can shift adoption from 0% to 100% at a fixed example count date — pass: originalPublished is 2026-10-28, which matches the v1 submission date on arXiv. The later feed date (Oct 31) is not used. > \[Submitted on 28 Oct 2026\] support — uncertain: All five claims are grounded in verbatim abstract text, and the limitations are appropriate. However, one sentence is not labeled as an inference: 'Their findings make the placement and recording of guidance central to whether agents follow existing patterns or new instructions.' It goes beyond the abstract's narrower claim that the findings motivate design-decision traceability. Smallest repair: label it as Escape Meridian interpretation, or tie it to the authors' own wording. > These findings motivate design-decision traceability that preserves not only implementation choices but also the rationale and priorities that should guide future changes. media — pass: The selected media is the verified, direct arXiv abstract page for this paper, which is the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/d62295b88bf9e67422f2491da367af1a25f231291435ab888f542d98ed0823ba/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/haichao-zhu-and-coauthors-ask-precedent-or-intent-d62295b8.json)