Quick answer

OpenAI released research on how developers are using Codex (the cloud-based autonomous coding agent, not the deprecated 2021 model). Three headline findings: 5× growth in active users in 12 months, multi-agent workflows are now normal, and developers routinely delegate tasks estimated at 8+ hours to AI. Agent-based software development is mainstream.

Codex is OpenAI's 2026 cloud-based coding agent. You assign it a GitHub task; it spins up a sandbox, reads your repo, writes code, runs tests, opens a PR. OpenAI just published usage research that quantifies how dev workflows are actually changing.

The three headline findings

  • 5× increase in active users over 12 months — much faster than ChatGPT's growth curve at the same maturity
  • Developers run multiple AI agents simultaneously — typical power user has 3-5 Codex tasks in flight at any time
  • Tasks estimated at 8+ hours of human work are routinely delegated to Codex — sometimes overnight, sometimes parallel during the day

What workflow looks like in 2026

The mental model has shifted from "AI helps me code" to "I manage AI agents that code." Developers describe their day as queue management — kick off 3-5 Codex tasks at the start of the morning, review the PRs over coffee, kick off the next batch. The IDE is no longer the centre of work; the agent queue is.

What kinds of tasks are being delegated

  • Refactors with clear scope (e.g., "extract this 500-line component into smaller pieces")
  • Bug fixes with reproducible test cases
  • Adding new endpoints to an existing API pattern
  • Writing test coverage for under-tested modules
  • Documentation updates across a codebase
  • Migration work (e.g., upgrade lib version, fix breaking changes)

What's NOT being delegated successfully

  • Greenfield architecture decisions
  • Open-ended performance optimisation
  • Anything that requires real-world user judgement
  • Tasks needing access to non-public context (internal Slack, design docs)
  • Tasks that fail tests in subtle, non-obvious ways

The "8-hour task delegation" number is striking but worth contextualising — it means tasks that would take a human 8 hours, not tasks the agent always completes well. Reliability on these tasks still varies; the time savings come from running them in parallel, not from each one succeeding.

What it means for hiring and dev productivity

  • Senior engineers are getting more productive without hiring
  • Junior engineering hiring at AI-native companies is way down
  • "How well do you manage AI agents" is now a real interview signal
  • Engineering org charts are flattening — fewer mid-level eng managers, more senior ICs
  • PR review is the new bottleneck — agents produce code faster than humans can review

Related reading

What Are Background Coding Agents? Best AI Coding Tools in Late 2026 From Chatbots to AI Agents

Bottom line

Agent-based dev is now mainstream. Codex usage 5×. Multi-agent workflows are normal. If you're still using AI as "smart autocomplete," you're behind the curve. The leverage is in delegating, not pair-programming.