Coding agents need explicit artifacts before they can earn trust

FeatX and MOA ask a large language model (LLM) to work through software evidence before editing code. FeatX maps repository features to Java classes and methods, then generates diffs through a three-stage agent. Its study reports a NASA-TLX workload drop from 12.5 to 7.4 against ChatGPT, plus 0.385 F1 for function-level modification localization on 38 commits.

MOA applies the same discipline to memory optimization. It turns profiling traces into validated anti-patterns, static checkers, and patches. On OpenHarmony, it reports 13 anti-patterns, 10,067 detected inefficiencies, 769 generated patches, and 92.5% expert acceptance. These results make the current emphasis clear: agents are strongest when the task surface is narrowed by source maps, runtime traces, and validation loops.

AdaTrans shows how compiler feedback can drive safer C-to-Rust migration. It maps Rust compiler errors to repair templates and documentation, then checks builds and output behavior. On 104 algorithmic problems, it reports 95.51% compilation pass rate, 81.09% solve rate, and 1.19% unsafe file rate.

JETO-Bench gives performance-fix agents a harder Java target. Its collection tool scanned 3,686 repositories and nearly 1.8 million commits, then produced 660 execution-time improvement patches and 91 manually verified executable tasks. OpenHands fixed 13 of those 91 tasks. The benchmark also finds that many Java projects lack tests that demonstrate execution-time improvement, which limits automatic scoring.

Agent capability is being measured with task cost and decision control in view. Artificial Analysis reports Claude Sonnet 5 at 53 on its Intelligence Index, 6 points above Sonnet 4.6. At max effort, it uses about 40% more output tokens per task and about three times the agentic turns on AA-Briefcase and GDPval-AA, raising measured cost to $2.29 per task at standard pricing.

Governance work adds a second constraint. A protocol study finds that MCP, A2A, ACP, ANP, and ERC-8004 do not fully support any of six governance dimensions: membership, deliberation, voting, dissent preservation, human escalation, and audit/replay. Serval’s Catalyst product points to a product-level answer for one setting: proposed automation changes are staged for review, and teams can require second-person approval before publication.

Two papers make code-agent evaluation less dependent on clean snippets. CoCoMUT extracts Java method context with callers, callees, class data, documentation, and source-bytecode reconciliation. Across 20 repositories, it emitted 56,512 method-context records and reconciled 97.8% of recognized project call targets to source methods.

The obfuscated-code study tests whether models still follow program behavior when names and control flow are distorted. Reasoning-tuned models track human task difficulty with positive Spearman correlations, while coder and instruction-tuned models show near-zero alignment. Control-flow flattening reduces accuracy as dispatcher complexity rises. This matters for review, audit, reverse engineering, and security work where surface cues may be misleading.