Solution
Braintrust uses Codex with GPT-5.5 as an agentic coding workflow around its observability and eval platform for AI products. Engineers paste live customer feature requests into Codex, point it at the relevant repository, and ask it to generate preview branches they can show back to customers in minutes instead of parking the request in backlog for later prioritization. For harder problems, the team writes a failing test, provisions a sandbox environment, and lets Codex run autonomously in that controlled workspace until it proposes a fix. This combines customer input, tests, and the existing codebase with a fast terminal-native model that can iterate without slowing down, so engineers can ideate with customers in real time, try more experiments, and move more quickly from request to working software.
Data flow
Customer feature requests are pasted into Codex alongside Braintrust's codebase context. For debugging or new ideas, engineers add a test and a sandbox environment, then GPT-5.5 iterates on code until Codex can produce a preview branch. The resulting branch is shown back to customers immediately, tightening the feedback loop before work enters the longer-term roadmap.