Developer workflow stack

Developers are converging on a split workflow: Linear holds the backlog, Claude plans the work, and Codex writes the code. (linear.app, anthropic.com, openai.com) Linear’s own docs describe the product as a system for product development, with issue workflows, triage queues, parent and sub-issues, and statuses that start in backlog by default. Those features make it a natural place to track bugs, quality-assurance checks, and handoffs before any model starts editing files. (linear.app, linear.app) Anthropic says Claude Code can read a codebase, edit files, run commands, and work across development tools, while Claude Opus 4.6 is tuned to “plan more carefully” and sustain long coding tasks in larger repositories. That makes Claude the planning layer in the stack developers are describing, especially for breaking a feature into smaller steps before implementation begins. (anthropic.com, anthropic.com) OpenAI describes Codex as a coding agent for software development that can write code, adapt to an existing project structure, and run multi-agent workflows with parallel worktrees and cloud environments. In that setup, Codex becomes the execution layer: a tool for drafting pull requests, refactors, and fixes after the task has already been scoped. (openai.com, openai.com) The pattern reflects a narrower use of artificial intelligence than the early “one prompt builds the app” pitch. Teams are separating product decisions from code generation, keeping architecture and prioritization in human-controlled systems while assigning models smaller, testable units of work. (linear.app, anthropic.com, openai.com) That division also lines up with how the tools are marketed in 2026. Linear says it is designed for the artificial-intelligence era, Anthropic markets Claude Code as an agentic coding tool, and OpenAI markets Codex as an “AI coding partner” built for real engineering work. (linear.app, anthropic.com, openai.com) The appeal is operational, not philosophical. A backlog tool records what should happen, a planning model turns that into a checklist, and an execution model handles the repetitive drafting that engineers can review, test, and reject if needed. (linear.app, anthropic.com, openai.com) The limits are visible in the product docs too. Claude Code depends on access to the repository and tools around it, and Codex only spawns new subagents when a user explicitly asks, which keeps the human operator in charge of scope. (anthropic.com, openai.com) So the emerging workflow is less “replace the engineer” than “formalize the handoff.” Planning lives in tickets and documents, code generation happens inside guardrails, and the final judgment stays with the team reviewing the work. (linear.app, openai.com, anthropic.com)