Braintrust Ships Customer Features in Minutes With OpenAI Codex

Key Takeaways

• Braintrust reduced feature request turnaround from backlog delays to under 10 minutes
• 50% of the engineering team switched to Codex within the first month
• Speed enables real-time customer iteration instead of async feedback loops

From Backlog to Branch in Minutes

Braintrust, the AI observability and evaluation platform, has changed how it handles customer feature requests. Instead of adding them to a backlog for later prioritization, engineers now paste requests directly into OpenAI's Codex, which generates working preview branches in minutes.

The company integrated Codex running GPT-5.5 into its development workflow via the Model Context Protocol, giving the AI deep access to its internal repository and experimental logs. The result: half the engineering team moved to Codex within a month.

For founder and CEO Ankur Goyal, the shift isn't just about writing code faster. It's about compressing the feedback loop with customers.

“The biggest change is not just faster coding. It's a faster feedback loop with customers.”

— Ankur Goyal, Founder and CEO of Braintrust

Real-Time Iteration Replaces Async Feedback

The old process was familiar to any software team. A customer requests a feature. It enters the backlog. Product managers prioritize it against other work. Engineers eventually build it. The customer sees the result weeks or months later.

Braintrust's new workflow collapses that timeline. Engineers copy a customer request into Codex, which creates a preview branch. The customer sees a working implementation in about 10 minutes, on average. This lets the team iterate with customers in real time rather than shipping something and hoping it matches what they wanted.

Goyal points to a specific technical advantage: Codex can output text in the terminal without slowing down. That sounds minor, but it changes how engineers interact with the tool.

“It sounds simple, but Codex can literally print more text in the terminal without getting slow, and other models just can't replicate that. The biggest gain is speed.”

— Ankur Goyal, Founder and CEO of Braintrust

Speed Changes the Experimentation Model

Goyal describes a shift in how he approaches problem-solving with AI tools. With slower models, he had to prompt step by step, guiding the model toward a specific solution. The overhead made experimentation expensive.

With Codex, he writes a test that demonstrates a problem, creates a sandbox environment, and lets Codex run. The speed makes this viable where it wasn't before.

A lead engineer at Braintrust, speaking anonymously, described the shift in architectural terms: "GPT-5.5's architectural shift toward agentic reasoning allows us to offload the entire 'feature-to-code' pipeline, not just code completion."

GPT-5.5's one-million-token context window helps here. The model can hold enough of the codebase in memory to understand architectural patterns and make changes that fit the existing system.

The Trade-Off Debate

Not everyone is convinced this workflow scales without consequences. On Hacker News, developers have debated whether rapid AI-generated code creates long-term maintenance problems. Some worry about "AI-generated technical debt" accumulating faster than teams can pay it down.

Others pointed to lighter concerns. The community jokingly referred to the "Goblin Fix," a patch that removed AI-generated mentions of goblins from terminal logs, as the most important GPT-5.5 update.

Braintrust's position as an AI observability platform may give it an advantage here. The company builds tools to evaluate AI outputs, which means it has infrastructure to catch problems that other teams might miss.

What This Means for Product Teams

The Braintrust case suggests a pattern worth watching. When AI code generation reaches a speed threshold, it changes more than developer productivity. It changes customer relationships.

Product teams have long talked about "shipping to learn." That usually meant weekly or biweekly releases with instrumentation to measure what users actually do. Braintrust's workflow compresses that cycle to hours, at least for certain feature types.

The approach won't work for everything. Complex architectural changes, security-critical code, and features requiring extensive testing still need traditional development cycles. But for customer-requested UI tweaks, workflow additions, and integration options, the speed advantage is real.

Frequently Asked Questions

What is OpenAI Codex with GPT-5.5?

Codex is OpenAI's code generation tool, now powered by GPT-5.5. It can write, modify, and debug code based on natural language instructions. GPT-5.5 adds a one-million-token context window and improved agentic reasoning for complex coding tasks.

How fast can Braintrust turn a feature request into working code?

Braintrust reports an average of about 10 minutes from receiving a customer feature request to generating a working preview branch. The demo video shows a 120-second example.

What is the Model Context Protocol used by Braintrust?

The Model Context Protocol (MCP) lets AI tools like Codex access internal repositories, logs, and development environments. This gives the AI deeper context about the codebase than simple copy-paste prompting.

Does AI-generated code create technical debt?

This is an active debate in the developer community. Rapid AI code generation can ship features faster but may create maintenance problems if the code isn't properly reviewed. Braintrust's observability tools help mitigate this risk.

Is this workflow suitable for all types of software development?

No. Braintrust's approach works best for customer-facing feature requests and iterative improvements. Security-critical code, complex architectural changes, and features requiring extensive testing still need traditional development cycles.

Source: OpenAI News