Coding AI

A study of engineering data from more than 700 software companies suggests that AI can increase code production without increasing software output. As generation speeds up, human review and downstream delivery may become the limiting factors.

ProgramGym introduces an environment-and-agent framework for long-horizon program reproduction, combining automatically repaired sandboxes with a structured development workflow.

A large study of software engineering activity finds that AI coding agents substantially increase code production, while longer and heavier human review absorbs much of the efficiency gain.

HERMES introduces Dev-Primitives, pairing repository artifacts with resident language models so components can reason about their own implementation, communicate, and support localized revisions. The work suggests that better agent harnesses can improve software engineering efficiency without relying only on larger models.

Most coding-agent benchmarks still focus on one codebase, while real software changes often span several related repositories. WideSWE introduces 120 real-world tasks to measure whether agents can plan, implement, and verify such coordinated changes.

Alibaba has open-sourced OpenCodeReview, a Go-based CLI that combines deterministic review workflows with LLM-driven code analysis. Its design aims to reduce omissions, unstable tool use, and token costs.

A new study addresses the uneven progress that can emerge when heterogeneous software-engineering tasks are trained in a shared reinforcement-learning pool. It develops category experts with an iterative Refresh–Repair–Expand loop, then consolidates them into one deployable model through label-routed multi-teacher on-policy distillation.

Xiaomi MiMo’s CodeMidas pipeline uses existing source code to discover functionality, generate executable tests, and filter coding tasks for reinforcement learning. The resulting tasks improved an agent’s performance across software repair, whole-program construction, and terminal benchmarks.

A new empirical study separates planning, action space, and context management in coding-agent harnesses. Its results show that the best design depends on model capability, task type, and available context budget.

Bend is a new programming language built around a simple premise: AI agents should not only generate code, but also prove that it respects declared rules. It combines Python-like syntax, native compilation, and a runtime intended for both CPUs and GPUs.

A private benchmark found no stable performance advantage for vendor-native harnesses when the underlying model was held constant. Task composition, timeout handling, and incomplete usage data proved equally important to the interpretation.

Moonshot AI has rolled out Kimi K2.8 Preview across Kimi Code and Kimi Work, positioning it as a lower-cost model with capabilities close to K3. The biggest product change is that 1M-token context is now available to every membership tier.

Cognition has introduced SWE-2, a coding model designed to improve not only task success but also the cost of reaching that result. Its benchmark profile is competitive with larger models, although performance varies considerably across evaluations.

Cognition, the company behind Devin, has raised $2 billion at a $48 billion valuation. The pricing suggests investors still expect several major AI coding companies to coexist rather than one winner taking the entire market.

A new study finds that code-repair models often rewrite more than necessary, even when their patches pass tests. Preservation prompts and reinforcement learning can improve the fidelity of their edits.

A large experiment finds that coding agents often disagree on third-party services. Repository language, project context and the way vendor information is presented can all change the final choice.

A new study presents an end-to-end specialization pipeline for competitive programming, combining curated problems, synthetic reasoning traces, post-training, and iterative test-time refinement. The reported system scored 535.4 out of 600 in an IOI 2026 evaluation.

PaperCompiler addresses a central weakness of paper-to-code systems: generated repositories can lose algorithmic details and consistency even when the code appears runnable. Its approach compiles implementation evidence from a paper into traceable, repository-wide specifications before code generation begins.

Harness-of-Harness is a framework built on existing coding-agent harnesses rather than a new code model. It organizes autonomous work into repeated planning, coding, testing, and evaluation cycles.

A dispute over one instruction-file format has become an enterprise tooling and governance issue. Shopify’s CEO considered banning Claude Code internally, while Anthropic’s response left developers unconvinced.

LinkedIn built a multi-agent review platform instead of simply placing a general-purpose AI reviewer in front of its repositories. The system combines independent reviewers, repository-specific rules, filtering, and production-grade operations to improve signal quality and measure whether developers actually adopt its findings.