Codex is no longer autocomplete. It's returning as an autonomous agent

From autocomplete to control layer

OpenAI announced a major overhaul of Codex. It is no longer a model for completing lines in an editor, but a long-running agent capable of planning, building, reviewing, and testing code across an entire codebase. The new version runs on the GPT-5.6 family of models (specifically Sol, Terra, and Luna) and is deployed in the Kiro development environment. The primary shift lies in the ability to maintain context over complex multi-step tasks and follow provided technical specifications and team standards. According to OpenAI's internal tests on the Terminal-Bench 2.1 benchmark, GPT-5.6 Terra in the Kiro environment demonstrates roughly an 82% cost-efficiency improvement compared to earlier solutions.

For development teams, it changes who approves the merge

For developers and product managers, this means a transition from AI writes a function for me to AI resolves a ticket and prepares a PR. Kiro takes a high-level intent, converts it into a technical design, and breaks it down into executable tasks. Crucially, it incorporates human review at key checkpoints a developer can pause the agent's work, review it, and guide it before the actual implementation happens. Furthermore, the model uses property-based testing to verify the functionality of its solution. This shifts the focus of a developer's work from writing boilerplate code to architecture and reviewing foreign (machine) work.

The cleanliness of the review queue will show if teams are serious

Marketing promises massive cost reductions and faster development. But the real limit will be the agent's ability to operate on large, poorly documented legacy systems. The demo handles the happy path over clean architecture, but as soon as Codex hits a dark corner of a ten-year-old codebase without tests, its spec-driven approach will hit a wall. Another risk is AI fatigue during approvals if the agent generates huge pull requests, humans will stop reading them thoroughly and start rubber-stamping them. This shifts responsibility but degrades quality.

Adoption outside the vendor's own story will decide

The proof that this generation of agents works will not be marketing percentages about speedups. It will be the tool's adoption in enterprise teams that don't have perfect technical specifications and where developers spend more time reading code than writing it. We will see whether developers prefer to keep control in their editor via Copilot, or if they delegate the entire layer to cloud agents.

Lilith's verdict

Long tasks will shift the burden from writing to auditing. Developers will become proofreaders of machine visions, and that transition will hurt.