Practical limits are now the main story in AI coding research

Token cost is the clearest practical constraint in this batch. The strongest evidence comes from a trajectory study on SWE-bench Verified with eight frontier models in OpenHands. Agentic coding used about 3500× more tokens than single-round code reasoning and about 1200× more than multi-turn code chat. Cost also swung hard on the same task, with runs differing by up to 30×. The paper ties expensive failures to repeated file viewing and editing, and it shows that models are poor at predicting their own token bill before execution. A second paper, R2Code, points to one concrete response: tighter retrieval and smaller context. It reports an average F1 gain of 7.4% on five traceability datasets while cutting token use by up to 41.7% through adaptive context control.

Repo-level generation is still weak when the task looks like real software work. RealBench builds 61 Python repositories from 20 domains, each with natural-language requirements, UML package and class diagrams, and human-verified tests. The best average Pass@1 is 19.39%. Repository size matters a lot: scores are above 40% below 500 lines of code and below 15% above 2000 lines. The benchmark also shows why small benchmark wins do not carry over cleanly. Only 44.73% of methods are standalone on average, and in the largest tier that drops to 26.23%, so dependency handling is the core problem. The paper also reports that whole-repo generation works better on smaller projects, while incremental generation works better once repositories get larger.

Verification and maintenance work remain a productive place for structured assistance. One paper on natural-language-to-Dafny generation reports near-universal failure with plain prompting, then much better results when the model gets method signatures and verifier feedback. On sampled tasks, Gemma 4-31B reaches 90.91% verification success and GPT-OSS 120B rises from 0% to 81.82% with signature-guided feedback. Another paper, R2Code, breaks requirements and code into aligned semantic parts, adds a consistency check, and reports gains on five datasets. Together these papers point to the same pattern: coding help improves when the model works against explicit structure, executable checks, and narrow evidence slices.

A smaller but concrete thread covers operational guardrails around agent access to real systems. QueryBear argues that database safety needs layered controls, not a single read-only rule. The stack it describes includes SQL parsing, table and column allowlists, AST-level query rewriting, pre-execution cost checks, statement timeouts, and audit logging. The examples are practical: blocking comment-hidden DELETE statements, pg_sleep(3600) denial-of-service queries, oversized joins, and joins that expose oauth_tokens. This is not a benchmark paper, but it fits the day’s emphasis on making agent actions reviewable before they touch production resources.