Bilingual

Senior SWE-Bench: open-source benchmark that assesses agents as senior engineers
24 days ago
- Senior SWE-Bench treats agents like senior engineers by using realistic, natural language instructions instead of over-specified requirements.
- It introduces a validation agent that writes behavioral tests to evaluate tasks, adapting to submitted solutions.
- Bug tasks are based on tricky user reports requiring runtime investigation, such as debugging logs and reproduction steps.
- Scoring combines runtime correctness tests with quality metrics to assess tasteful code solutions.
- Tasks are sourced from PRs in diverse repositories and involve multi-phase, multi-stack features or bugs with significant runtime investigation.
- Instructions are naturally under-specified, with a median length 31% that of SWE-Bench Pro.
- Feature tasks can span multiple services, averaging 11 files touched per task, and are long-horizon, requiring hundreds of steps.
- A leaderboard shows top-performing models like Claude Opus 4.8 achieving a 24.0% solve rate, with frontier models failing over 75% of the time.
Most rewrites serve the engineer, not the business
24 days ago
- Rewrites often prioritize the engineer's preferences over business value, as shown by an example of porting a CakePHP application to Laravel without tangible benefits.
- Existing code holds historical knowledge and bug fixes; rewriting it risks reintroducing past issues and losing valuable 'scar tissue' documented by Joel Spolsky.
- Unfamiliarity with a codebase or framework can lead to misjudgments; assuming previous developers made reasoned choices is crucial before deeming code 'broken'.
- Legitimate reasons for rewrites include end-of-life runtimes, loss of key personnel, high feature development costs, or business needs that outgrow the current design.
- AI can accelerate code generation but may omit context for past fixes, leading to rediscovery of old bugs and perpetuating a cycle of rewrites without addressing root causes.
- To break the rewrite cycle, document the rationale behind code decisions clearly, ensuring future engineers or AI models understand the 'why' and not just the 'what'.
Model Training as Code
a month ago
- Model training complexity requires specialized teams and manual coordination is inefficient.
- Aleph Alpha's Savanna implements Model Training as Code (MTaC) for hermetic, one-click training runs.
- Manual training leads to human errors, lost learnings, and fragmented team ownership.
- MTaC offers composability, consensus via version control, and provenance through code history.
- Savanna uses trunk-based development, CI for validation, and immutable artefact versioning.
- It enables hyperparameter sweeps with caching to avoid redundant computations.
- The system runs on Flyte in Kubernetes, with artefact tracking and real-time monitoring.
- MTaC improves iteration speed, facilitates team restructuring, and supports auto-research.
LLM's as a Different Kind of Intelligence
21 days ago
- LLMs represent a different kind of intelligence, distinct from human intelligence.
- Human analogies like 'junior engineer' or 'abstraction layer' are flawed because LLMs lack intent and deterministic reasoning.
- LLMs combine vast knowledge with an inability to understand intent, leading to subtle drift from user goals over time.
- Probabilistic behavior from LLMs challenges the deterministic foundations of software engineering.
- Effective use requires learning LLMs' unique failure modes rather than forcing familiar comparisons.
Zero-defects code: the prescient Microsoft memo from 1989
20 days ago
- Microsoft's 1989 Zero Defects memo advocated for principles like continuous integration and test-driven development long before they became mainstream.
- The memo highlighted problems in Microsoft's culture, such as rewarding developers for checking in incomplete code and treating schedules as more important than quality.
- Chris Mason argued that writing perfect, zero-defects code was achievable and emphasized the need for a mindset shift to prioritize quality over speed.
- Key solutions proposed included having a working, shippable product daily, fixing bugs immediately, and shifting testing to quality assurance rather than bug-finding.
- The memo suggested practical techniques like code reviews, writing tests before debugging, and using asserts to verify assumptions, aiming to reduce bugs and improve predictability.
- Leadership committed to supporting realistic schedules and pushing back against pressures to cut corners, emphasizing that taking minor heat early was better than major issues later.
Does Code Cleanliness Affect Coding Agents?
20 days ago
- The study investigates whether the cleanliness (structural and stylistic quality) of code affects the performance of autonomous coding agents.
- It introduces a minimal-pair evaluation protocol, comparing repositories identical in architecture, dependencies, and behavior but differing in static-analysis violations and cognitive complexity.
- Results from 660 trials with Claude Code show that code cleanliness does not significantly change task pass rates.
- However, agents working on cleaner code use 7–8% fewer tokens and reduce file revisitations by 34%, indicating improved efficiency and reduced computational cost.
- The findings highlight that traditional maintainability principles remain relevant in AI-driven development, influencing agent behavior alongside model choice, harness, and prompting.
Interview: Drew DeVault on an AI-free version of Vim
18 days ago
- Drew DeVault forked Vim as 'Vim Classic' to create an AI-free version, maintained entirely by humans without LLM-assisted code.
- DeVault criticizes generative AI for its environmental impact, including increased energy consumption, pollution, and strain on supply chains, disproportionately affecting poorer communities.
- He argues AI causes 'deskilling', making users less competent, and subverts software engineering by prioritizing shortcuts over careful planning and craftsmanship.
- Ethically, DeVault highlights AI's role in human rights abuses, such as in Gaza and Iran, and its support for fascist regimes through companies like Palantir and Anthropic.
- Politically, he views AI as a tool of oppressors that harms the working class and accelerates fascism, with tech companies often operating unethically, similar to Uber's model.
- DeVault expresses despair over the lack of consensus in the tech industry against AI but finds hope in building community and continuing to advocate for human-centric approaches.
- The interviewer shares similar concerns as a creative professional, noting isolation in fields embracing AI, and finds encouragement in DeVault's stance.
SWE-1.7 Reach Near GPT 5.5 and Opus Intelligence
17 days ago
- Cognition launches SWE-1.7, a frontier-level AI model for software engineering at lower cost, advancing cost-performance.
- SWE-1.7 achieves 42.3% on FrontierCode 1.1, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual.
- Training improvements include preserving entropy via top-p sampling replay, multi-cluster infrastructure for scalability, and fault tolerance.
- Self-compaction enables long-horizon tasks by summarizing working state, with alternating length penalty to balance efficiency.
- High-quality data curation ensures task difficulty and prevents cheating, enhancing model alignment and trustworthiness.
- SWE-1.7 shows condensed chain-of-thought reasoning and thorough codebase exploration compared to base model Kimi K2.7 Code.
Grok 4.5
17 days ago
- Grok 4.5 is released in collaboration with SpaceXAI as the most intelligent model, designed for tasks beyond software engineering.
- It can handle difficult, long-running tasks in various fields like software engineering, data science, finance, and legal work by creatively using tools.
- Cursor subscription plans include significant model usage with double usage for the first week and added safeguards for cybersecurity.
- Grok 4.5 is a mixture-of-experts model trained with trillions of tokens from Cursor data, learning from software and developer-agent interactions.
- Training data was broadened to include high-quality STEM tasks, research papers, and knowledge work for proficiency across domains.
- Reinforcement learning was applied in realistic environments to teach problem investigation, tool usage, error recovery, and result verification.
- A distributed agent system scaled environment construction, using previous models to accelerate progress on difficult problems.
- Grok 4.5 is available on multiple platforms with pricing at $2/M input tokens and $6/M output tokens, and a fast variant at higher rates.
- Grok 4.5 and Composer 2.5 will both be supported, with future models planned for the Composer size.
Interview with Mitchell Hashimoto about Ghostty and Zig
16 days ago
- Mitchell Hashimoto, creator of various DevOps tools and currently working on Ghostty and Vouch, discusses his projects, terminal development, and open-source philosophy.
- He explains that interviews with him often focus on his transition from software engineering to business founder or on his products like HashiCorp tools and Ghostty.
- Hashimoto started Ghostty to improve his technical skills, explore GPU programming, desktop systems programming, and learn Zig, leading to a fast, feature-rich, cross-platform terminal emulator.
- He believes terminals should be optimized for text-based applications with clear security and composability, rather than pushing them to extremes like full application platforms.
- Hashimoto advocates for new terminal protocols, such as an n-screen API for multiple screens and a button protocol for interactive elements in scrollback history.
- He emphasizes that open-source maintainers have no obligation to users, but he strives to improve software while balancing user feedback with personal vision and avoiding scope creep.
- Hashimoto supports forking projects and believes the open-source community should embrace more personal forks, rather than expecting polished, product-like support from maintainers.
- He praises Zig for its continuous improvement and willingness to make breaking changes, and notes AI can ease the pain of such changes by automating updates.
- Hashimoto's library design approach involves learning from various ecosystems and porting enjoyable concepts, as seen in his focus on development-specific 'nouns' in Vagrant.
- He is critical of overly complex tech stacks and industry trends but accepts mainstream technologies for practical reasons like community and hiring.
- Hashimoto's principles for projects like Ghostty are personal reflections, leading to feature-rich, cross-platform core with native GUI, attracting like-minded contributors.
- He values diverse programming language cultures and respects Zig's unique stance, despite not agreeing with all aspects, while expressing dislike for Rust's community culture.
- To ensure quality, Hashimoto uses his own software extensively and employs AI for prototyping, but emphasizes understanding user needs and shipping code with care and empathy.
- He recommends learning how computers work through low-level systems programming in languages like C, Zig, or Rust, and studying standard library implementations to grasp underlying concepts.
The Lindy Effect in Software
18 days ago
- The Lindy effect suggests that the longer something has existed, the longer it is expected to continue existing.
- In software, this means older technologies are often more robust and reliable than newer ones.
- Choosing established tech offers benefits like stability, mature ecosystems, predictable performance, industry acceptance, and reduced risk.
- Examples include C language and SQL, which have endured for decades and remain relevant.
- For software engineers, applying the Lindy effect involves prudent adoption, sticking to proven foundations, planning for longevity, and embracing evolutionary improvements.
Why Write Code in 2026
14 days ago
- Software engineers build and maintain the infrastructure (software factory) that enables agents to make changes and ship quickly.
- Writing code remains valuable for direct thinking in the execution environment, enhancing attention, understanding, and ownership of system architecture.
- Agents are not compilers but more like interns; they require precise context, and human involvement helps prevent fragility and slop in code.
- Hands-on coding helps identify and fix weaknesses, establish architectural principles, and improve the software factory's robustness.
- Details matter in software factories, from architectural patterns to algorithms; agents encourage evaluation and CI, but human intervention is essential for addressing weak spots.
Perfectly Hitting the Wrong Target: The Story of an AI Code Review Benchmark
13 days ago
- Benchmarks are often perceived as authoritative and objective but may lack depth if not scrutinized.
- The author critiques AI Code Review Bench methodology, arguing it lacks clear problem definition and splits AI code review into two problems: human assistance and machine verification.
- Human-focused code review prioritizes recommendations to optimize limited attention, while machine-focused code review emphasizes exhaustive analysis for repair agents.
- The paper acknowledges challenges like reviewer imperfections and Goodhart's Law but may overemphasize proxies like agreement with human reviewers over outcomes like reducing production failures.
- Organizational context and severity of issues are overlooked, with universal benchmarks potentially averaging different notions of software quality.
- Future code review systems should separate human comprehension (interface and recommendation design) from machine verification (exhaustive hardening and repair).
- Despite criticisms, the paper advances discussion with contributions like acknowledging human reviewer limitations and proposing metrics beyond recall and precision.
Agent Harness Engineering
13 days ago
- Agent harness engineering focuses on the scaffolding built around AI models to create effective coding agents.
- A harness includes system prompts, tools, infrastructure, orchestration logic, hooks, and observability components.
- Harness engineering treats agent mistakes as signals to improve the system, adding constraints to prevent recurrence.
- Key harness components include filesystem and Git for state, bash and code execution as general-purpose tools, and sandboxes for safe execution.
- Memory and search mechanisms enable continual learning, while context management techniques combat context rot.
- Long-horizon execution is supported through loops like Ralph Loops, planning, verification, and planner/generator/evaluator splits.
- Hooks enforce rules and provide feedback, while AGENTS.md serves as a concise rulebook injected into prompts.
- Harnesses evolve with models; as models improve, harness components shift to address new capabilities and failure modes.
- The model-harness training loop shows that models become co-trained with specific harnesses, affecting performance.
- Harness-as-a-Service frameworks simplify agent development by providing pre-built runtime components for customization.
- Industry trends show convergence in harness patterns across top coding agents, indicating shared best practices.
- Future directions include multi-agent orchestration, self-improving harnesses, and dynamic tool assembly for tasks.
In defense of not understanding your codebase
13 days ago
- The article argues that in large software systems, having only a partial understanding of the codebase is acceptable and often necessary.
- It critiques Peter Naur's 'Programming as Theory Building' by stating that rebuilding large systems from scratch is impractical due to complexity and that abandoned systems can be revived through gradual learning.
- The author suggests that engineers in large teams must work with imperfect mental models, making educated guesses and dealing with consequences.
- Trade-offs are discussed, noting that factors like team turnover, legal requirements, and dependencies can hinder maintaining a full theory of the code.
- The article concludes that while maintaining a complete theory is desirable, in professional settings, other values like speed or compliance may take precedence.
A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI
12 days ago
- The study examines the adoption, retention, and impact of command-line AI coding agents like Claude Code and Copilot CLI at Microsoft in early 2026.
- Findings indicate that initial adoption spread mainly through social networks among engineers.
- Retention was linked more to engineers' coding activity levels rather than demographic factors.
- Adopters merged approximately 24% more pull requests compared to their baseline, using merged PRs as a proxy for output.
- The productivity lift persisted over a four-month observation period.
- The results suggest CLI coding agents are not uniformly adopted or mere novelties, and peer influence is crucial for rollout strategies.
- Organizational token costs can be high, making accurate assessment of adoption and impact essential to avoid wasted investment.
What are Forward Deployed Engineers, and why are they so in demand? (2025)
12 days ago
- Forward Deployed Engineers (FDEs) are software engineers who alternate between being embedded with customer teams and core product engineering teams, performing a mix of software, sales, and platform engineering.
- Palantir pioneered the role in the early 2010s, originally calling it 'Delta', and until 2016 had more FDEs than standard software engineers, significantly shaping the role.
- FDEs help close sales by assisting customers with product integration, often traveling to customer sites (e.g., 25-50% onsite time) and working in varied environments like factory floors.
- They contribute to core product development, adding features based on customer needs, and often work with ambiguity to build solutions, such as AI integrations.
- OpenAI differentiates FDEs from Solutions Architects, with FDEs being more hands-on, writing code on customer infrastructure, and aligning with research goals to advance product roadmaps.
- Ramp employs FDEs in pods following four core principles, emphasizing collaboration with sales, embedding in product teams, and driving product engineering decisions.
- The role is in high demand in 2025, especially for integrating LLMs and AI solutions, as FDEs bridge gaps between technology and customer adoption in enterprises.
- FDEs are distinct from similar roles like Solutions Engineers or Agent Engineers by their dual focus on customer success and product contribution.
- Companies like Palantir use FDEs to overcome bureaucracy in traditional enterprises, enabling faster software integration and customer value delivery.
A Large-Scale Empirical Study of AI-Generated Code in Real-World Repositories
12 days ago
- The study conducts a large-scale measurement comparing AI-generated and human-written code in real-world repositories.
- It addresses limitations of prior lab-based studies with synthetic benchmarks and small-scale tasks.
- Metrics cover code-level aspects like complexity, style, security, and commit-level traits such as size, frequency, and stability.
- Findings reveal small differences in code-level metrics between AI and human code in real-world settings.
- Insights include variance in security quality across programming languages and new data on code duplication and commit characteristics.
- The results provide implications for AI-assisted programming practices.
Ten Steps Towards Happiness (2015)
11 days ago
- Pieter Hintjens acknowledges donations for his children's future and thanks contributors.
- He has participated in various podcasts, interviews, and video talks, including on Software Engineering Daily, Ruby Rogues, and multiple conferences.
- Keynote talks were given at events like DomCode 2015, Coding Serbia, EuroPython, and Tech Mesh.
- He invites contact for potential speaking engagements at events or organizations.
C Strings: A 50-Year Mistake – By Trần Thành Long
10 days ago
- Null-terminated strings in C are outdated and prone to errors; length-based strings are superior.
- Length-based strings store both pointer and size, enabling efficient operations like substrings without extra allocations.
- C strings' ambiguity in length definitions (e.g., snprintf) leads to bugs and confusion in handling null terminators.
- Length-based strings support arbitrary binary data, including null bytes, unlike C strings which rely on sentinels.
- C strings force unnecessary allocations for common operations, while length-based strings allow efficient slicing and referencing.
- Modern hardware and compilers reduce performance benefits of sentinel-based strings; length-based designs are more robust and clear.

software engineering

Senior SWE-Bench: open-source benchmark that assesses agents as senior engineers

Most rewrites serve the engineer, not the business

Model Training as Code

LLM's as a Different Kind of Intelligence

Zero-defects code: the prescient Microsoft memo from 1989

Does Code Cleanliness Affect Coding Agents?

Interview: Drew DeVault on an AI-free version of Vim

SWE-1.7 Reach Near GPT 5.5 and Opus Intelligence

Grok 4.5

Interview with Mitchell Hashimoto about Ghostty and Zig

The Lindy Effect in Software

Why Write Code in 2026

Perfectly Hitting the Wrong Target: The Story of an AI Code Review Benchmark

Agent Harness Engineering

In defense of not understanding your codebase

A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI

What are Forward Deployed Engineers, and why are they so in demand? (2025)

A Large-Scale Empirical Study of AI-Generated Code in Real-World Repositories

Ten Steps Towards Happiness (2015)

C Strings: A 50-Year Mistake – By Trần Thành Long