Goldman limits Claude Code and Copilot; a git queue for agent work
Goldman Sachs limits Claude Code and GitHub Copilot to engineers, one developer details a git-based queue for agent work, and another explains why he will not use LLMs. Briefs cover arXiv studies and tools.
Top stories
Goldman Sachs restricts Claude Code and GitHub Copilot to its engineers, cutting off traders and bankers
- Goldman Sachs updated internal guidance in October to limit Anthropic's Claude Code and Microsoft's GitHub Copilot to its engineers, Bloomberg reported, after a review of productivity and risk concerns.
- Traders had used the tools to monitor client flows and model positions; strats keep access, and the bank has about 12,000 engineers among roughly 47,000 staff.
- State of AI says agent-built code needs a named engineer owner before touching production data, and that existing information barriers and approval rules still apply.
Developer describes git-based ticket queue to cut context switching across Claude Code and Codex sessions
- Sergey Kuhne describes a method for 'AI brain fry' from juggling many Claude Code or Codex sessions, organizing work as a priority queue of markdown tickets stored in git.
- An agent researches and builds tickets tagged safe, never deploys to production, and a script sends a prioritized owner inbox, read through a generated Claude artifact page.
- He closes each item by returning it to the queue, which he says lets him focus on one thing at a time; the remaining challenge is tooling.
Developer Kristian Polso says he will stop using LLMs and AI agents in professional coding
- Kristian Polso, a developer with about 10 years of machine learning experience, wrote that he will not use LLMs or AI agents in his professional work.
- His reasons include cognitive impact, verbose and poorly fitting generated code, maintainability, enjoying writing code, datacenter environmental cost, and GDPR and CLOUD Act concerns with US-based providers.
- He concedes AI can give a workable plan when he is stuck, but says the help is not worth it, and he does not aim to change others' views.
In brief
- Lovable shares lessons from 100,000 pull requests, including deleting most of its agent code. Lovable CEO Anton Osika says the team learned from more than 100,000 pull requests, with users creating over 70 million projects, and rethought how much freedom to give its agent model. (lovable.dev)
- Recto launches free visual canvas for editing what Claude Code, Codex, Cursor and OpenCode build. Recto is a free desktop tool that runs locally, detects installed coding agents, and lets users see every change, edit elements directly or reprompt while building UI, SVG, 3D and motion. (recto.design)
- arXiv study finds coding agents follow stale instruction files and none repaired them. A paper on arXiv analyzing 732 trajectories across 22 maintenance tasks finds agents follow stale guidance; Claude Sonnet 5.5 needed about two extra tool calls on a stale test command, and none fixed the file. (arxiv.org)
- arXiv paper finds 469 GitHub Security Advisories with coding-agent involvement across 189 repositories. A paper on arXiv mining GitHub Security Advisories finds 469 candidate advisories, 448 with direct attribution, with Claude the most frequent tool, followed by Copilot and Cursor. (arxiv.org)
- Tracybot links agent-generated code changes to the prompts that produced them. An arXiv paper presents Tracybot, an open-source tool for VS Code and Git that maps line-level agent changes to Tasklets; a 7-person focus group and 24-person survey found it useful for code review. (arxiv.org)
- Ben Page praises Amp's orbs, cloud virtual machines for coding agent threads. Ben Page writes that Amp's orbs run each agent thread in a virtual machine that keeps running when the laptop closes, and says he has stopped local development. (ben.page)
- Carette argues AI use is not a choice between full delegation and rejection. Carette uses a runner-and-robot analogy to propose a relay model, where engineers keep the hard parts and AI handles well-defined segments, and critiques DHH's report of 37signals defaulting to agents. (carette.xyz)
Chosen by software from the day's news, written by AI from the pages it links to, and checked against them by a second model. Published by Rapid Review, Inc.