Coding-agent research is centering on evidence, harnesses, and bounded compute

Automated program repair (APR) work is using execution evidence as first-class input. PracRepair combines static code context with failing-test traces, variable values, branch outcomes, validation diagnostics, and trace diffs. With GPT-4o, it reports 162 correct fixes on Defects4J V1.2 and 171 on V2.0, plus many fixes not found by ReInFix.

A separate code-correction study tests the same idea in a simpler loop: generate code, run it, return compiler errors or failing test details, and revise. GPT-o4-mini leads pass@1 on the 450-problem core set, while the study finds that syntax and runtime failures are easier to repair than logic and algorithm errors.

Test generation remains a weak point. In 86,156 agent-authored test-file patches, 80.2% contain weak or no explicit oracle signals. The finding matters because a test file can execute code and still check no expected behavior.

Benchmarking work is naming the measured object more carefully. One position paper argues that coding-agent scores include the model, harness, tools, environment, task setup, and verifier. Its cited Terminal-Bench example shows a fixed Claude Opus 4.6 scoring 79.8% ± 1.6 with ForgeCode and 58.0% ± 2.9 with Claude Code, a 21.8-point spread across harnesses.

The agent-skills evaluation paper turns that concern into a measurement procedure. It generates tasks and hidden rubrics from each skill, runs agents with and without the skill, and scores instruction following and goal completion. Across about 500 skills, 1,000 tasks, 19 agent-model configurations, and 38,000 valid trajectories, relevant skills add 5.5 to 22 points depending on the model.

LoopCoder-v2 gives a concrete compute-scaling result for code models. The 7B Parallel Loop Transformer reuses shared blocks across loops and keeps cache cost down with shared-key-value attention. In the reported suite, two loops improve the average score from 38.0 to 46.5. On SWE-bench Verified, the score rises from 43.0 to 64.4, while three and four loops fall sharply.

Operational writing in the corpus applies a similar bounded-systems view to agent workstations. Long-running agents need isolated computers, separate ports, state, dependencies, and shell access. The practical claim is that quality depends on specs, tests, type checks, linting, observability, and final pull-request review, not on continuous human editing of generated code.