Reading up on Devin

The article argues that as AI models internalize agent harness capabilities, the remaining harness evolves into an interface for managing scarce human attention rather than directing model behavior.

Products that are hard to evaluate programmatically are likely hard for users to verify, so design for verification first.

As AI models commoditize measurable tasks, lasting value lies in 'untrainable' work requiring private data, trust, organizational change, and domain-specific authority.

Cognition's FrontierCode benchmark measures code mergeability, finding even top models like Claude Opus 4.8 score only 13.4% on its hardest 50 tasks.

Cognition CEO Scott Wu argues AI coding agents like Devin will expand software creation rather than eliminate engineering jobs despite industry layoffs and competition.

Devin 2.2 adds end-to-end desktop testing, self-review autofix, and 3x faster startup so the agent can autonomously plan, code, test, and fix code.

Cognition uses its own Devin agents across Slack, Linear, and CLI to merge 659 PRs weekly, automating code review, bug triage, and design-system audits.

Devin is an AI software engineer that autonomously codes, debugs, and deploys applications from natural language prompts.

COBOL developers find AI coding tools helpful for documentation and simple tasks but not yet a threat due to legacy system complexity and compliance constraints.

A developer demonstrates how to build a custom async coding agent by combining Slack, Modal, GitHub, and Claude Code/Codex headless mode, arguing this makes cloud coding agents like Devin undifferentiated.