[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85676-en":3,"doc-seo-85676-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},85676,549758252649,"Ivy","https://ap-avatar.wpscdn.com/avatar/8000253669c5317157?_k=1778319167496531819",8,"Research & Report","What Context Does a Coding Agent Actually Need to Act","A study examines what context a modern coding agent truly needs to edit code effectively. Focusing on the find/act boundary, the work holds localization fixed via an oracle and varies only code representation, scoring results on SWE-bench Verified. Natural-language summaries provide minimal behavioral signal compared with source code, and adding surrounding file context yields little improvement. Compressed context matches full files at far lower token cost. The paper also releases a rigorously validated instrumentation suite and finds T=0 non-determinism can flip outcomes.","arXiv :2607 .09691v1 [ cs .LG] 19 Jun 2026  \nWhat Context Does a Coding Agent Actually Need to  \nAct?  \nBrian Sam-Bodden∗  \nIntegrallis Software  \n[bsbodden@integrallis.com](bsbodden@integrallis.com)  \nAbstract  \nA modern coding agent can hold an entire repository in its context window. Most of its reading is wasted—and the interesting question is not how much context an agent can use, but what it actually needs. We study that question at the moment it matters most: when the agent must edit code. Separating finding the work site from acting on it, we hold localization fixed with an oracle, vary only how the code is represented, and score context against real issue resolution on SWEbench Verified. The answer is starkly minimal. The signal lives in the code being edited itself: natural-language summaries of it answer almost none of the behavioral questions that the source answers (4/45 vs. 27/45, held-out repositories, independent judge), and the gap belongs to the representation, not the summarizer  \n—a frontier model’s summaries score exactly as poorly as a 3B model’s. The surrounding context hardly matters either: across every multi-file instance in Verified, under a protocol frozen before any data, rendering a file’s remainder as UML skeletons and signatures resolves no more issues than deleting that remainder outright (N=70, exact McNemar p=0 .75) . That was our registered hypothesis, and it failed. Compressed context, meanwhile, matches whole files at a third of the tokens: a resolved issue costs 19K context tokens, not 94K. The instrument also  \nyielded a finding the field should keep: temperature-0 API inference flips ∼9% of per-instance outcomes between byte-identical runs. That is a noise floor under every small effect reported on this benchmark, including ours. We release the instrument—gold-validated environments, per-instance proof that every reference edit is expressible from every arm’s context, deterministic patch construction, and pre-registered hypotheses whose nulls we publish.  \n1 Introduction  \nWhen a software engineer fixes a bug, she doesn’t re-read the repository. She finds the few functions that matter, reads them closely, and lets the rest of the codebase fall into a vague map of what exists where. Coding agents are built as if the opposite were true: context windows now hold entire repositories, and the natural instinct is to fill them. That instinct has a price—tokens, latency, dollars—and surprisingly thin evidence behind it. Frontier agents already resolve a large fraction of SWE-bench Verified [5], yet ContextBench [7] finds that sophisticated retrieval scaffolds often barely beat trivial baselines, with a wide gap between context explored and context used. So the question worth asking is not whether we can fit the code, but: for a given code task, what is the minimal context an agent actually needs—and in what representation?  \nWe separate two sub-questions that are often conflated:  \n• Find—locating where the work happens;  \n∗ Code and data: [https://github.com/integrallis/act-context](https://github.com/integrallis/act-context)  \nPreprint.  \nbehav ioral-probe accuracy (Wilson 95%  \n0.8  \n0.6  \n0.4  \n0.2  \n0.0  \n27/45  \nfll1 , 4utrcoke  \nsum tkritten by frontier model  \nsum tkritten by 3B model  \nsi~g2n1te + docstring  \na frontier summarizer scores identically to a 3B one: the gap belongs to the representation, not the summarizer  \nFigure 1: The find/act boundary. Behavioral-probe accuracy by context representation (n=45:  \n15 held-out classes from seaborn/pylint/pytest × 3 probes minted from each class’s test file; the same model—Claude Sonnet 4.6—reads the context and answers in every arm, so only the representation varies; answers graded by an independent-family judge, κ=0 .84 against a second judge) . The frontier/3B labels indicate the summarizer that wrote the representation, not the answerer. Source answers 60%; natural-language summaries answer ∼9%—identically for a frontier and a 3B summ","cbCaituphTLlj5Yv","https://ap.wps.com/l/cbCaituphTLlj5Yv","pdf",406887,1,9,"English","en",105,"# Abstract\n# Introduction\n## Find—locating where the work happens\n## Act—knowing enough to edit correctly\n# Contributions","[{\"question\":\"What key question does the paper address about coding agents’ context usage?\",\"answer\":\"It asks what minimal context a coding agent actually needs for the moment it matters most: editing code, rather than how much context can fit in a context window.\"},{\"question\":\"How do the experiments separate finding the work location from acting on it?\",\"answer\":\"They fix localization using an oracle and vary only how the code is represented, then score resolution on SWE-bench Verified.\"},{\"question\":\"What are the main findings about natural-language summaries versus source-code context?\",\"answer\":\"Natural-language summaries answer far fewer behavioral questions than source code, and the gap is attributed to representation rather than the summarizer used.\"}]",1784205534,23,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"what-context-does-a-coding-agent-actually-need-to-act","",{"@graph":35,"@context":84},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/what-context-does-a-coding-agent-actually-need-to-act/85676/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What key question does the paper address about coding agents’ context usage?","Question",{"text":74,"@type":75},"It asks what minimal context a coding agent actually needs for the moment it matters most: editing code, rather than how much context can fit in a context window.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How do the experiments separate finding the work location from acting on it?",{"text":79,"@type":75},"They fix localization using an oracle and vary only how the code is represented, then score resolution on SWE-bench Verified.",{"name":81,"@type":72,"acceptedAnswer":82},"What are the main findings about natural-language summaries versus source-code context?",{"text":83,"@type":75},"Natural-language summaries answer far fewer behavioral questions than source code, and the gap is attributed to representation rather than the summarizer used.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,126,129,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":123,"show_sort_weight":124,"slug":125},"Religion & Spirituality",20,"religion-spirituality",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":124,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]