All articles
84 published

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.

Development
AI Lets Developers Do Work They Can't Check
A Microsoft survey of 331 developers finds AI's biggest unlock is work beyond their own skill, which is the work they are least equipped to check. That needs its own verification.

Product Design
Synthetic Users Match the Headline and Get the Segments Wrong
Two October 2026 studies find synthetic research panels reproduce the overall answer but distort differences between groups, in opposite directions. The headline check can't catch either.

Development
Coding Agent Debt Is Broken Behavior, Not Messy Code
Two 2026 studies find coding agents struggle on code other agents wrote, yet maintainability metrics don't explain it. Earlier tests that already fail do.

Design Engineering
Figma Specs Lets the Coding Agent Rewrite the Spec It Builds From
Figma's Specs beta lets coding agents read and update the spec over MCP. That can turn a statement of what to build into a record of what got built, with no documented check.

Product Design
Engage-to-Unlock: Making Users Write First Cost No Extra Time
A 1 October 2026 ETH Zurich and Google DeepMind experiment locked AI drafting until users wrote their own argument. Total time held flat, and the AI wrote less of the result.

Development
Coding Agents Write More Code, but Reviewers Keep Sending It Back
A Harvard study of 718 firms finds coding agents raise code output about 30% without raising finished work, and the review data points at draft quality, not volume.

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.

Product Design
AI Models Beat Humans on the Design Brief, and Their Screens Look Alike
An 8 October 2026 benchmark of 15 AI models found all passed over 90% of brief criteria, yet their UI designs resembled each other more than human work. Brief-based review cannot see that.

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.

Development
Developers Trust AI When It's Easy to Check, and Code That Runs Passes
Stack Overflow's 2026 survey shows developers trust AI only when they can easily validate it. Three September studies suggest the easiest check, running the code, misses what AI gets wrong.

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.