code agents
TrialTechniques
AI systems that generate code and submit changes from short specifications.
Why it's here
Placed in Trial: 10 article(s) of evidence from 5 source(s), led by research-stage coverage, with 4 in the last 30 days. Confidence 84%.
Evidence (10)
- 5Martin Fowler·9/29/2026regulationFragments on AI Agents, Persistence, and Regulation
This post reflects on how coding agents can be powerful but also difficult to control, arguing that effective use requires strong discipline and specialized know-how. It highlights concerns about agentic systems becoming overly persistent or unsafe, and suggests that responsibility should rest with the organizations that train and deploy them.
- 2Simon Willison·9/23/2026framework_updateSan Francisco meetup on agentic engineering
Simon Willison and Jesse Vincent are hosting an evening Birds of a Feather session in San Francisco on October 14 for people building with coding agents. The event is framed as an informal show-and-tell focused on early-stage ideas, unusual experiments, and unfinished projects rather than product pitches.
- 7The New Stack·9/22/2026securityAI coding agents need secret-safe context boundaries
The article argues that AI coding agents can unintentionally expose secrets because they ingest broad local context such as source files, logs, and environment data. It says traditional security checkpoints like code review and CI are no longer enough, and organizations need controls that prevent secrets from crossing from local development environments into model context or external provider systems.
- 5Hacker News·9/8/2026open_sourceI-have-ADHD: A prompt skill to keep coding agents concise
The project introduces a prompt skill called “i-have-ADHD” for coding agents, designed to prevent them from burying the answer in long explanations. It appears as an open-source GitHub repository and has drawn significant Hacker News discussion around improving agent responses for developers.
- 7Hacker News·8/18/2026researchLLMs Make Benchmark Overfitting Easier
The post argues that coding agents and LLMs can now game performance benchmarks much more easily, producing apparent gains that do not translate to real-world performance. It uses a regex engine experiment as an example, showing strong benchmark results alongside poor holdout performance and algorithmic regressions.
- 5The New Stack·8/9/2026researchCoding agents can be evaluated by their work
The article argues that coding agents are difficult, but not impossible, to evaluate, and that they should be assessed as whole systems rather than as standalone models. It proposes judging them by executable outcomes, change quality, trajectory, human intervention, cost, and production impact instead of relying only on pass/fail test scores.
- 5Hacker News·7/23/2026researchWhy Software Factories Fail
This Hacker News item highlights an essay arguing that software factories do not succeed by relying on engineering harnesses alone. The piece appears to focus on advanced context engineering for coding agents and the limits of trying to industrialize software creation without addressing broader workflow and system constraints.
- 4Simon Willison·7/20/2026researchCoding agents make reverse engineering cheaper
The article argues that coding agents have lowered the cost of reverse-engineering and automating home devices. Because it is now easier and cheaper to write, test, and discard code, the ROI calculus for building and maintaining these fragile automations has changed.
- 4The New Stack·7/16/2026researchThe Real Bottleneck in Independent Deployments Is Validation
The article argues that many platform teams already have strong deployment tooling, but still ship changes in batches because they cannot validate each change independently with enough confidence. It says shared environments and release trains were practical responses to scarce validation capacity, and that coding agents may worsen the problem by increasing change volume and batch size.
- 6Hugging Face Blog·4/16/2026open_sourceHugging Face adds agent-assisted PR tooling for MLX model ports
Hugging Face describes a Skill and test harness that help port language models from transformers to mlx-lm, with the goal of making new models available in MLX soon after they land in transformers. The post frames the tools as aids for contributors and reviewers rather than full automation, and argues that open source needs new workflows as code agents increase PR volume.