Tech News
← Home  ·  All topics

Ai Coding Agents

14 GoKawiil briefs on this topic · RSS for ai coding agents

Harvard study: AI coding agents move bottleneck to code review

Harvard researchers studied software teams using AI coding agents and found that while the agents sped up code generation, the slowdown point shifted to human code review. Reviewers now spend more time vetting AI-generated code, offsetting some of the speed gains from automated coding.

Claude Code agent found leaking credentials into public GitHub repos

Researchers found that a command used by Anthropic's Claude Code AI coding agent inadvertently saves files containing credentials and other sensitive data, which then end up committed to public GitHub repositories. The issue reportedly has exposed thousands of secrets, creating security exposure for organizations using the tool.

Concept proposes storing software tasks as Markdown files in Git repositories

A workflow called 'Tasks as Code' suggests replacing external trackers like Jira or Redmine with Markdown task files stored directly in a code repository's /docs-style folder. The approach extends the existing 'Docs as Code' practice, letting AI coding agents read tasks, code and documentation in one place and record updates through ordinary Git commits. It is linked to Spec-Driven Development, where Markdown specifications define intended behavior before an agent implements it.

Microsoft releases Scope, open-source tool for testing AI agent compatibility

Microsoft has released Scope, an open-source platform built on Kubernetes that lets development teams run repeatable experiments to see how AI coding agents interact with their SDKs, CLIs, MCP servers and documentation. The tool is available under the MIT License and is designed to turn observed agent behavior into actionable fixes for product teams. Microsoft cites the Azure Cosmos DB team, which used Scope's evaluations to resolve 20 Agent Kit issues and 27 documentation issues.

Malicious AGENTS.md files found tricking AI coding agents into running attacker commands

Forever Security researchers discovered a malicious AGENTS.md instruction file in a customer's Downloads folder that caused an AI coding agent to attempt installing a rogue npm package called 'function-color.' They reproduced the attack against OpenAI Codex and showed a similar technique against Claude Code that could leak OAuth tokens and hijack accounts, all using only crafted text instructions placed in the agent's working directory.

Essay argues AI coding agents reshape programmer's role, not eliminate it

A developer essay sketches a hypothetical 2031 workday in which AI agents complete most coding tasks overnight, opening pull requests with tests and recorded demos, while leaving ambiguous business decisions for a human to resolve. The piece draws a parallel to earlier predictions that compilers would make programming obsolete, arguing that each layer of automation shifts rather than removes human judgment. The author frames this as a plausible near-future extension of present-day agentic coding tools rather than a confirmed roadmap.

LayerFS introduces auto-checkpointing filesystem for AI coding agents

LayerFS is a filesystem designed for AI coding agents that automatically creates a checkpoint at every step an agent takes. Each file operation is recorded as a commit, and each Bash command commits once it finishes executing. Calls that make no changes do not generate a commit, avoiding unnecessary history noise.

Study of 700+ firms finds AI coding agents boost code volume, not shipped software

A study covering more than 700 software companies found that AI coding agents raised lines of code by 30 percent, commits by 20 percent, and pull requests by 23 percent, but did not meaningfully increase the rate at which firms resolved tracked features or issues in tools like Jira. Researchers attribute the gap to code review, which took 49 percent longer on average after AI tools were adopted, with more revisions required and more reviewer comments per pull request.

Study: AI coding agents boost code volume but not software output

Harvard researchers Fiona Chen and James Stratton analyzed Jellyfish analytics covering 300 million work events across more than 700 software firms and 700,000 employees from 2021 to March 2026. They found that while AI coding assistants and agents increase the volume of code generated, firms show little evidence of increased software output or reduced employment, because human code review becomes a bottleneck, with longer reviews, more revision requests, and more reviewer comments.

Gitframes launches as npm package combining Photoshop, After Effects and Blender tools for AI coding agents

Gitframes is a new npm package that bundles compositing, motion graphics, typography and 3D scene rendering into a single lightweight tool meant to be driven by code rather than desktop software. It lets an AI coding agent write a TypeScript composition and render an MP4 directly on GPU hardware via WebGPU, Dawn, Metal or Vulkan, without using a headless browser or screenshot pipeline. The project is currently in beta, with APIs and some features still subject to change.

Factory CEO fires board advisor Chris Degnan, who joins rival Cognition as CRO

Factory co-founder Matan Grinberg said on X that he removed Chris Degnan from his role as board advisor after alleging Degnan shared confidential Factory information with competitor Cognition. Hours later, Degnan announced on X and LinkedIn that he had joined Cognition as chief revenue officer. Grinberg said Degnan had previously denied any interest in joining Cognition before disclosing ongoing talks with the company this week.

New open-source tool Corral guarantees full cleanup of agent-spawned processes

A developer released Corral, a Linux command-line tool that runs a shell command with a time limit and ensures every process it spawned, including detached daemons, is terminated before returning. It uses cgroup v2's cgroup.kill for guaranteed cleanup when available, or acts as a child subreaper tracking processes via /proc and pidfds otherwise, and exits with code 120 if it cannot verify a clean kill. The tool requires Linux 5.11 or later, with 5.14 needed for its enforced cgroup mode, and is not intended as a security sandbox.