FurtherAI's CTO says his team now merges 5x the code each day, at the same revert rate

Sashank Gondala is the cofounder and CTO at FurtherAI. He posted a 5-minute read on X about nine months of changes to the team's engineering process. The headline number: the code changes each engineer merges a day grew about 5x since January. The revert rate stayed under half a percent. The post went to 159 points on Hacker News.

Gondala writes that coding agents alone did not get them there. In his words, "Every time we sped up one step, work piled up at the next."

So the team walked the pipeline in order. Each step gets a month and a number.

The pipeline, with the numbers

Agents write code now, close to 100 percent of the time. Tests come with it: 91 percent of changes include tests, up from 15 percent. AI reviews 99 percent of changes before a person does, up from 60 percent.

CI runs on only the tests touched by a change, in parallel. That is 20x the tests in the same 17 minutes.

Bug reports go through a bot called ClaudeTag. Anyone writes a bug in Slack. ClaudeTag figures out duplicates, files a ticket, and routes it to the right person. It now files 60 percent of all tickets.

The step Gondala highlights hardest is rules. Sometimes a bug or a review comment exposed a lesson only one engineer carried. The team turned it into a small Python check. The check runs on every commit and every pull request, fails the build, and prints the fix. The team has 21 such checks now, up from zero. Agents hit them, correct themselves, and move on.

Up to a third of all changes now merge with no person involved. Low-risk changes that pass tests, the automated checks, and AI review merge on their own.

What the extra speed cost

Gondala is careful about the price. Through the summer, much of the team's work went into tests and tooling. Customer-facing changes per engineer stayed between 17 and 30 a month. In September, that number jumped to 45. October is on pace for 48.

He also names the new problem. The team got much faster at writing code. Deciding what to build did not get faster. In his worked example, a feature used to take ten days of plan, build and review. If building becomes 3.5x faster but planning and review stay the same, the feature as a whole gets 1.75x faster. Planning and review are now 70 percent of the time, up from 40 percent.

The bet, he writes, is to push more ownership to engineers rather than adding management. He also names a shift for engineering managers. They should "be in the work" and try features on staging as a user would. Deciding what to build matters more as writing it gets cheaper.

A side effect he notes: non-engineers started merging code too. The team's designer has merged about 50 changes since August.

Source

Sashank Gondala on X: Building an AI-native engineering org, posted while serving as cofounder and CTO at FurtherAI.