At JADEV, we currently use Codex and Claude Code as complementary tools in our engineering team. We previously worked with Cursor, GitHub Copilot and Gemini with Antigravity. That history reflects our needs and successive evaluations, not a universal ranking of those products.
We reassess our choice every quarter. A useful tool on one project may be unsuitable on another, particularly when access or confidentiality requirements change.
What the client is buying
The client buys software that meets a need, verifiable delivery and an accountable team. They are not buying a volume of generated code. Architecture decisions, scope tradeoffs and acceptance of a change remain human responsibilities.
AI can help investigate an existing codebase, prepare a change, explore test cases or review a proposal. Its answers are working material to verify. A convincing answer, or agreement between two assistants, is not evidence that a behavior is correct.
Using Codex and Claude Code together
We select the tool and context for the task. A proposal developed with one assistant can be discussed or reviewed with the other when project rules allow it. That adds another perspective, but does not replace engineer review or relevant checks.
Tasks need boundaries: the problem, expected behavior, affected files and permitted access. A bug fix is not blanket permission to rewrite an application. Repository rules and client conventions belong in the working context.
Our quarterly review and tool benchmark
Each quarter we review the tools against their usefulness in our work. Our evaluation criteria emphasize correctness, instruction following, work on existing systems, the quality of proposed tests and the human effort needed to correct the result.
We also consider total time to an acceptable change, usage cost and data handling conditions. A fast response that requires extensive rework is not a gain.
Meaningful comparisons require comparable tasks, recorded versions and configurations, and keeping failures in the assessment. Benchmark examples should be public, synthetic or explicitly permitted for that use. We publish no unmeasured scores or productivity gains here. Codex and Claude Code are our current choice, not a promise that this combination will remain the best fit for every engagement.
Privacy starts with the project's rules
Assistant use must respect client commitments. Before providing context, establish which data is permitted, which account and configuration are appropriate, and which provider terms apply.
Our working rules are straightforward: minimize context, exclude secrets and credentials, prefer synthetic or anonymized data, and do not send confidential code to an unauthorized service. Logs, screenshots and tickets can contain sensitive information too.
A tool's brand alone does not establish its retention, training or data location policy. Those depend on the service, agreement and configuration selected. We therefore make no blanket promise of zero retention or exclusively European hosting. If a client rule prohibits a use, that use is excluded from the project.
Verify before delivery
An AI-assisted change is still a software change: understand the diff, check added dependencies and permissions, run appropriate tests and inspect the result. A test that repeats the implementation without checking the requirement can be green and useless.
For sensitive features, review must cover authorization, exposed data, failures and rollback options. Production actions and contractual approvals do not become autonomous because an assistant suggests them.
What we can clarify before starting
Together we can define permitted tools, usable data, necessary access and approval points. The level of assistance depends on your constraints. We remain accountable for the delivered work.
Taking over software or clarifying risk before a build? Explore our architecture and delivery audit. For AI in your own product, see our AI engineering service.