Coding agents are learning to find the fault, not just write the fix
New work on log-guided localization and version-bound agent loops points at the same gap: agents that ship code need proof about what broke, and gates on what they can merge.
- 49% of real-world production issue reports do not name any of the files that the eventual fix changed. Self-reported m:…9771
- INQUEST identifies the changed files in production issue reports with 68.6% top-1 accuracy. Self-reported m:…9771
- CLEW reaches 77.3% Acc@1 and 81.3% MRR on the LogLoc-Issues benchmark, and cuts wall-clock time by 73% compared with LocAgent. Self-reported m:…9764
- Staged agentic decision-making beats free-loop approaches by 14 percentage points in naming faulty functions. Self-reported m:…9771
- Only 21% of organizations have a mature governance model for autonomous agents. Self-reported m:…4060
- 51% of IT and AI executives named managing and monitoring AI agents at scale as their top operational challenge. Self-reported m:…4060
The bug report doesn't say where it hurts
Production bug reports are vague about the one thing a fixer needs most. In a preprint on log-guided root-cause localization, 49% of real-world production issue reports do not name any of the files that the eventual fix changed. The report describes a symptom; the fault lives somewhere else, and the agent has to find it. m:…9771
Two preprints attack that gap with logs. INQUEST, a staged agent for log-guided root-cause localization, achieves 68.6% top-1 accuracy for identifying changed files in production issue reports. CLEW, a log-anchored backward tracing method, achieves 77.3% Acc@1 and 81.3% MRR on the LogLoc-Issues benchmark, maintains 97.6% Acc@1 on the LogLoc-Airflow subset, and reduces wall-clock time by 73% compared to LocAgent. m:…9771 m:…9764
Structure beats a free loop
The design lesson is structure. Staged agentic decision-making outperforms free-loop agentic approaches in naming faulty functions by 14 percentage points, according to the INQUEST preprint: the agent that works through stages beats the agent left to loop freely. Loop Engineering, a preprint on coordinating production coding agents, makes the same point at the system level. It uses six specific elements, trigger, context projection, authority boundary, verifier, stop condition and write-back rule, to define agent loops, and version-bound test locks can effectively prevent merges when new commits invalidate prior approval and evidence. m:…9771 m:…9770
A systematic mapping of LLMs in DevSecOps finds the field moving the same way: current LLM-based security systems are shifting from single-shot prompting to structured context, iterative reasoning, fine-tuning, and agentic workflows. But the mapping also finds the gaps. Research focuses far more on the Code and Test stages than on Deploy and Review, most systems operate in isolation within specific stages, and comparison across systems is hindered by heterogeneous tasks, datasets, metrics, and baselines. m:…9772
Finding is not fixing
Finding the fault and fixing it are different jobs, and the fixing side is further behind. A preprint on moving from vulnerability discovery to fixes reports that iterative LLM pipelines with multi-stage feedback loops improve the quality of automated software vulnerability repair compared to single-pass generation, and it proposes LLM4CVE as a standardized benchmark for evaluating LLM-based vulnerability repair agents. The proposal itself is an admission that the field still lacks a shared way to say whether a fix is any good. m:…1705
Running them is the hard part
The operational side is where most organizations are stuck. Dynatrace's account of operating AI agents reports that only 21% of organizations have a mature governance model for autonomous agents, and that 51% of IT and AI executives named managing and monitoring AI agents at scale as their top operational challenge. The same account says 45% of IT and AI executives lack clear rules for when agents can act autonomously, and 42% have limited real-time visibility to trace and troubleshoot agent behavior. Conventional monitoring cannot show whether an agent used the right context or tool, followed policy, or produced an outcome worth the cost, Dynatrace says. m:…4060
The pattern across these papers is that the agent's job is splitting in two: diagnosis, where logs and staged reasoning are producing measurable results, and authority, where version-bound locks and governance models decide what the agent is allowed to do. The first half is improving fast; the second half is where the risk sits, and it is the half most teams have not built yet. m:…9770 m:…4060
What is still unresolved
- INQUEST and CLEW are evaluated on benchmarks; neither preprint shows results from a live production rollout.
- There is no shared benchmark result yet for whether the fixes these agents propose actually hold up in production.
- The governance figures come from a vendor's account of its own survey, not an independent study.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.