Tag

Claude Code

A bold vintage screen-print illustration in parchment, burgundy and olive of a large faceted ascending bar chart topped with a flag, beside a big burgundy circle.

Development

A Better Coding Model Didn't Get Developers More Work Merged

A 17 November 2026 Harvard paper uses a forced Claude Fable 5 shutdown to show that higher benchmark scores did not raise developers' merged work, which argues for buying on workflow, not leaderboards.

A bold vintage screen-print illustration in stone, deep indigo and clay of a geometric signpost with two arrow boards pointing in opposite directions beside a large indigo circle.

Development

Coding Agents Flag the Stale Rule, Then Follow It Anyway

Two October 2026 preprints show coding agents naming a problem or claiming to stop, then acting against their own words. The diff and a written rule, not the reply, are the levers.

A bold vintage screen-print illustration in cream, deep green and vermilion red of a large faceted rubber approval stamp pressing down on a sheet of paper with a signature line.

Development

Agent Approvals Logged as the Tech Lead's: Bashi AI's Merge Audit

A 19 October 2026 paper audits a fintech startup where 27% of merges rested on approvals a Claude Code session filed under the tech lead's name. The fix is identity, not better review.

A bold vintage screen-print illustration in denim blue, butter cream and coral of a large striped boom barrier arm lowered across the frame beside a small curling paper note.

Development

Claude Code Memory Reminds; Only Runtime Checks Enforce

Writing a correction into CLAUDE.md or auto-memory does not reliably change a coding agent's behavior. One operator's logs and a June paper point to hooks, which enforce but over-block.

A bold vintage screen-print illustration in cream, olive green and brick red of a printed code listing with a large red checkmark and a round approval stamp.

Development

AI Code Reviewers Can Coach Attackers to Approval

A 4 October 2026 paper shows an AI pull-request reviewer's own comments guide an attacker to approval with the exploit intact, just as GitHub makes Copilot's approval count.

A bold vintage screen-print illustration in pale sage, deep ink blue and vermilion of a large rubber approval stamp pressing a checkmark mark onto a form sheet.

Creative Tooling

Claude Code Mods Can Approve Commands Before the Classifier Sees Them

Anthropic's new Claude Code mods can approve a tool call after rules and hooks decide, skipping the auto-mode classifier. The real safety layer is now plugin vetting, and solo developers carry it.

A bold vintage screen-print illustration in cream, deep teal and ochre of a bulletin board with two paper printouts, a before and an after sheet, each held by a round push pin.

Development

PixelLeak: The Screenshot Request That Leaked Internal UIs

Glow's PixelLeak report found 13,000+ internal screenshots in public GitHub repos. The cause was a routine review request, and the fix sits in your agent's skill files.

A bold vintage screen-print illustration in pale cream, deep pine green and rust orange of a large calibration dial gauge crossed by a socket wrench, set against a scoreboard-tile band and diagonal stripes.

Development

The SWE-Bench Leaderboard Can No Longer Tell Models Apart

Two September 2026 papers, days apart, find SWE-bench Verified can't statistically separate its top coding agents, and the harness deciding the score resets with every model swap.

A bold vintage screen-print illustration in terracotta, dark ink and warm ochre of a long paper receipt unrolling from a geometric spool, torn off after only a short filled-in stub while the rest of the ribbon runs blank.

Development

The Coding Agent's Self-Report Covers One Action in Eleven

A 5,851-session study finds coding agents' self-written wrap-ups cite about one action in eleven, and drift back toward the plan exactly when it was abandoned.

A bold vintage screen-print illustration in kraft-gold, deep aubergine and muted sage green of a tall stack of cardboard shipping parcels bound in crossing packing tape, the topmost parcel bearing a large round provenance seal-label still fully intact and unopened.

Development

AI Coding Assistants Skip the Labels Before They Install

A pre-registered audit of 1,920 trials finds AI coding assistants open a provenance signal before installing 0.5% of the time, and never once verify one.

A bold vintage screen-print illustration in cream, deep indigo navy and dusty teal of a large two-button computer mouse centered between blocky choropleth-map quadrants, with a tiny closed chat bubble icon left unused in the corner.

Product Design

The Prompt Box Lost 94% of the Time to the Ordinary Mouse

An OOPSLA 2026 study put a chat box next to click-and-drag controls in the same editor and found users typed prompts for only 6% of their edits.

A bold vintage screen-print illustration in cream, deep slate-blue and red of an open instruction manual, its ruled checklist page intact with red-pen checkmarks while its facing map page shatters into scattered geometric fragments.

Development

AGENTS.md Works as a Rulebook and Fails as a Tour

ETH Zurich tested 138 AGENTbench instances plus 300 SWE-bench Lite cases and found AGENTS.md's instructions change agent behavior, but the repo overview doesn't, while adding 20% to run cost.