Introducing Governor, a context-driven model router for coding agents Bito Ai

How PubMatic cut coding agent costs 64% with Governor

Redwood City, California

Summary

PubMatic cut its coding agent costs by 64% after routing its agents through Bito’s Governor, and its agents now finish tasks 34% faster. PubMatic is a publicly traded ad tech company (NASDAQ: PUBM) with about 450 engineers who use coding agents like Cursor and Claude Code every day. As agent use grew and PubMatic started building a software factory, its AI model bill climbed fast. Governor brought that bill down without a drop in the quality of the work.

64%

Lower coding agent spend

34%

Faster agent tasks

450

Engineers

500+

Repositories

The Challenge 

PubMatic’s engineers use Cursor and Claude Code across a leading advertising infrastructure platform that handles about 840 billion ad impressions a day and spans more than 500 repositories. Claude Opus is their most preferred model. As the engineers handed more of their work to agents, PubMatic’s spending on AI models went up sharply.

Then PubMatic began building a software factory. In a software factory, agents take on whole pieces of work from start to finish instead of helping one engineer at a time. Agent spend started climbing much faster than headcount. Most requests went to Opus by default, even small tasks that a cheaper model handles just as well. Opus is also one of the most expensive models, and its cost was putting real strain on engineering budgets.

The team needed to get costs under control while keeping the quality of the agents’ work where it was.

“We want every engineer using agents, and we want to build a lot more of our software that way. At the rate our model bill was growing, we were going to hit a ceiling. We needed to spend less on each task without our engineers seeing a drop in quality.”

– Mukul Kumar, President, Engineering, PubMatic

The solution

Governor sits between PubMatic’s coding agents and the AI models they call. PubMatic already ran its agent traffic through a gateway, so Governor plugged in behind it and engineers kept working in the same tools.

Governor cuts cost in two ways. It hands the model the right code at the start of a task, so the agent spends far fewer tokens opening files and searching for what it needs. It also sends each request to a model that fits the job, so routine work stops going to the most expensive model.

Customers pick how Governor makes that trade. Quality mode favors the strongest models, Cost mode pushes hardest on savings, and Balanced mode sits in between. PubMatic tried all three and got the best results by mixing them by role:

  • Junior engineers run in Cost mode. Their tasks tend to be smaller and well defined, and the lower-cost setup handles them well.
  • Senior engineers run in Balanced mode. They take on harder design and cross-service work, where a stronger model is worth paying for

Moving a group from one mode to another is a setting, so PubMatic keeps adjusting as it learns which work needs what.

Results

PubMatic’s coding agent spend dropped 64% across the engineering org. Quality held up, and engineers didn’t have to change how they work.

PubMatic checked quality with its own benchmarks before rolling Governor out. Governor brought costs down sharply with almost no change in the quality of the work, and in every one of PubMatic’s tests it scored well above Claude Sonnet. When PubMatic tested Quality mode, the results matched Opus.

Agent tasks also got faster. Because Governor gives the agent the code it needs up front, the agent skips most of the slow back and forth of opening files and searching. Tasks now finish 34% faster. Engineers spend less time waiting on their agents and get through more tasks in a day.

The savings also changed what PubMatic can afford to build. Money that would have gone to model costs now goes into the software factory.

“Governor has been really helpful. It’s made our agent work so much more efficient that we can invest even more in our agentic SDLC efforts.”

– Mukul Kumar, President, Engineering, PubMatic

About PubMatic 

PubMatic (pubmatic.com) is an AI-powered ad tech company headquartered in Redwood City, California, and listed on NASDAQ as PUBM. It processed nearly 350 trillion ad impressions in 2025 and runs its own infrastructure, which it has been building for more than 15 years. Its engineering organization has about 450 engineers.