Codex is no longer just autocomplete. It now runs as an autonomous agent for long tasks

OpenAI is turning Codex into an agent

In its recent updates, OpenAI has quietly shifted from treating Codex as a better autocomplete to a full-fledged agent runtime. Instead of completing a single block of code, the system can now take instructions at the feature ticket level, plan the approach, break it down across files, and edit them iteratively. This model is much closer to how tools like Devin operate.

The boundary of delegation shifts for engineering teams

This shift represents a fundamental change in daily engineering workflows. If a model can independently handle tasks spanning multiple files or functions, the senior developer's focus moves from writing boilerplate to approving and orchestrating architecture. Instead of “write me a function,” the instruction becomes “here is the issue, solve it and open a pull request.”

Demos handle the happy path, production bugs do not

The ability to break a ticket into steps is one thing, but the reality of production code full of legacy dependencies is another. While the system excels in greenfield projects or well-isolated issues, the agentic loop often hallucinates or gets stuck when it hits non-standard internal conventions or deep build-process errors. The system's capacity to evaluate its own failure states remains its weakest link.

The code review queue will prove its viability

The proof of success for this approach won't be the volume of generated lines, but the throughput of the review process. If the agent generates hundreds of lines that a senior developer must meticulously analyze and rewrite, net productivity will actually drop. The signal for adoption will be whether these agentic pull requests pass the CI/CD pipeline and human review as smoothly as those from junior team members.

Lilith's verdict

The senior engineer just handed repository access to an intern who learns at lightning speed but still can't recognize a burning server.