AI coding assistants

Evidence (18)

• 3Hacker News·10/6/2026research
  Why Vibecoding Feels Less Fun Than Hand-Coding

  The article argues that AI-assisted vibecoding can be exciting because it quickly turns ideas into working prototypes, but it frontloads the thrill rather than sustaining the deeper satisfaction of building software. The author says AI is useful for projects that would otherwise not get built, but it often does not deliver the long-term enjoyment of hard-earned accomplishment, learning, and careful craftsmanship.

• 6Hacker News·10/3/2026product_launch
  Default hard budget caps are becoming essential

  The article argues that pay-by-usage services should default to hard monthly budget caps that shut usage off once a limit is reached, rather than merely sending warnings. It cites new spend-limit features from AWS and Google Cloud as signs that this approach is becoming more common, especially as coding agents make it easier to spin up services that can incur unexpected charges.

• 7InfoQ·9/30/2026research
  CodeScene Case Study Shows Agent Refactoring at Scale

  CodeScene published a case study describing coding agents refactoring 300,000 lines of C in three weeks at an estimated cost of about $4,000 in tokens. The result was verified with a frame-by-frame replay harness, but practitioners have questioned the scope of the work, the metric used, and how much the harness influenced the outcome.

• 7The New Stack·9/26/2026security
  Agent security shifts from access controls to action controls

  The article argues that identity and access management are necessary but insufficient for agent security because agents can route around perimeter controls by choosing alternate paths. It proposes inside-out controls at the agent harness level, where each action is checked before execution against policy, authority, and target system.

• 7The New Stack·9/19/2026breakthrough
  Engineer Says Verification Enables 2,000 Production PRs a Month

  An engineer at SpaceXAI described a personal agent workflow that reportedly helped ship about 2,000 pull requests a month to production. The article argues that the key enabler is automated verification, backed by a rich runtime and debugging environment that lets agents test and inspect changes end to end.

• 4Hacker News·9/18/2026research
  Empirical study examines harness design for coding agents

  This arXiv paper studies how harness design affects the performance and evaluation of coding agents. It compares design choices empirically to identify which setups lead to more reliable results. The work is mainly relevant to developers and researchers building agentic coding systems.

• 6The New Stack·9/18/2026research
  AI code review is overloading senior engineers

  The article argues that heavy AI-assisted coding is shifting senior engineers from writing code to reviewing large volumes of generated pull requests, creating review bottlenecks and burnout. It says AI output is often plausible but flawed, and recommends codifying recurring review feedback into rules and tests to reduce manual review load.

• 5InfoQ·9/10/2026research
  When Spec-Driven Development Pays Off

  The article argues that AI coding assistants have shifted software development bottlenecks from code generation to code verification. It highlights how spec-driven development can help detect and mitigate cases where AI-generated behavior diverges from the intended requirements, while also reducing security weaknesses and recurring bug patterns.

• 8Simon Willison·9/6/2026research
  OpenAI’s research acceleration and internal coding agents

  The article describes how OpenAI’s research and engineering work increasingly relies on coding agents, with agentic engineering reportedly accelerating in 2026. It also discusses recursive self-improvement (RSI) and suggests a notable rise in AI spend per researcher in late July, possibly tied to internal access to a model later released as GPT-6 Astra.

• 6Hacker News·9/6/2026research
  OpenAI reports how coding agents are speeding research

  OpenAI shares early internal data showing how coding agents are being used to support AI research. The post highlights changes in experiment speed, task complexity, and overall research acceleration as agent adoption grows.

• 5The New Stack·8/31/2026framework_update
  Why coding agents need a context development lifecycle

  The article argues that skills, prompt instructions, agent configurations, and rules files should be treated as software artifacts with their own development lifecycle. It proposes a Context Development Lifecycle (CDLC) with four phases: generate, evaluate, distribute, and observe, to improve reliability and reduce regressions in coding agents.

• 5The New Stack·8/26/2026research
  AI agent traces are becoming application data

  The article argues that execution traces from AI coding agents are no longer just diagnostics when teams need to store, review, and reuse them as durable records of work. It explains that these traces can include model calls, tool use, file reads, command results, and state transitions, which creates access-control and retention requirements similar to other product data.

• 4Simon Willison·8/26/2026research
  Paul Dix on AI-Assisted Software Refinement

  Paul Dix argues that AI can generate very large codebases and then iteratively refine them into reliable software when paired with a verification system and clear direction. He cites an example of AI producing and improving around 1 million lines of code for software now running on millions of developer machines. The remarks highlight confidence in coding agents and AI-assisted programming for complex software development.

• 2Simon Willison·8/22/2026research
  Beyond Code Review for Coding Agents

  Simon Willison argues that the key to using coding agents productively is giving precise instructions and then verifying the result reliably. He notes that checking every line of generated code is not always the best way to validate software changes.

• 7The New Stack·8/21/2026research
  New Benchmark Targets Large-Scale Code Refactoring

  Researchers from several institutions introduced SWE-Bench ProMax, a multilingual benchmark focused on large-scale code refactoring across seven programming languages. The benchmark includes heavily curated real-world commit tasks and shows that current frontier AI coding agents still struggle, with the best model resolving only 41.2% of cases.

• 2Hacker News·8/14/2026research
  HN post criticizes oversized AI-generated pull requests

  A Hacker News post argues that AI-assisted development is leading to oversized pull requests that are harder for humans to review and maintain. The author says small PRs are meant to improve reviewability and asks developers to avoid shipping large, hard-to-digest diffs, even when AI can generate them quickly.

• 4Hacker News·8/12/2026research
  Build Wide, Ship Narrow

  The article argues that AI makes it cheaper to build features, reshape designs, and split finished work into small pull requests, changing the traditional need to pre-plan code boundaries. It recommends planning the product and design first, then building on a wide branch, demoing early, and only later decomposing the code into narrow PRs based on the structure the implementation reveals.

• 4The New Stack·8/8/2026research
  AI Adoption vs. AI Usage in Engineering Teams

  The article argues that common AI adoption metrics, such as seat activations, token spend, and weekly active usage, mostly measure usage rather than real workflow change. It says lasting adoption only happens when AI becomes part of team norms or recurring processes, not just individual experimentation.