How we Construct with AI

That is half one in all a three-part collection on how HubSpot reworked with AI. Half two covers how we develop with Agent-first GTM. Half three is how we function as an AI-first firm.

All the pieces we construct at HubSpot exists to assist our prospects develop. So when generative AI emerged, our engineering staff didn’t simply see a productiveness software; we noticed a possibility to construct higher merchandise and get extra worth into prospects’ palms sooner.

And when off-the-shelf AI instruments hit their ceiling, we didn’t simply search for higher ones. We constructed the platform beneath them. That call compounded quicker than we anticipated. As a result of all of our AI is constructed on a shared basis, each new functionality we ship makes the entire system extra highly effective and prospects get a extra constant expertise throughout all the pieces they use.

Right this moment, we’re capable of innovate at a tempo that merely wasn’t potential earlier than. 100% of our engineers use AI, and we’ve seen a 73% improve in strains of code written by our engineers.

We didn’t get right here in a single day. It took three phases, actual infrastructure funding, and a willingness to construct what didn’t exist but. Right here’s how we did it.

Section 1: Productiveness with Co-pilots (2023-2024)

In 2023, giant language fashions had simply crossed the edge of being genuinely helpful in a coding context. One of the best resolution for utilizing AI in engineering was to start out with what was confirmed. At the moment, it was code completion: a human writes code, and AI copilots counsel what comes subsequent.

We rolled out a coding copilot and acquired to 30% adoption shortly. Then we pulled the incident knowledge, in contrast groups utilizing the copilot in opposition to groups that weren’t, and proved AI adoption didn’t negatively impression the reliability of the product.

With that knowledge in hand, we eliminated the guardrails and gave everybody copilot entry. Adoption shot previous 50% in a single day. This taught us a lesson in how we make selections. Measure, show, then scale.

By the tip of Section 1, 80% of engineers have been utilizing AI instruments. We noticed a 51% enchancment in engineering velocity, that means engineers have been transport working code to manufacturing considerably quicker, and a 7% improve in strains of code up to date per engineer. We proved AI might make each engineer quicker with out compromising product reliability.

Section 2: Scaling with Coding Brokers (2024-Mid 2025)

The following step was autonomous coding with brokers. Our groups might immediate the instruments to finish end-to-end duties. The brokers might learn context, write code, run assessments, and repair errors, all whereas the engineer reviewed and steered. We felt strongly this was the way forward for engineering and dedicated totally.

The actual constraint got here shortly. Off-the-shelf coding brokers couldn’t entry inner construct techniques, our libraries, or confirm that code really labored in the environment. So, we constructed these agent integrations ourselves utilizing MCP, a normal that enables AI brokers to connect with exterior instruments and techniques, and deployed them to each engineer. To drive adoption, we organized occasions to offer engineers devoted house to be taught, experiment, and construct confidence with new instruments. Agent utilization went from zero to 80% adoption in a month.

The following problem was scale. Engineers wished a number of brokers working in parallel, in a single day, with out supervision. So we constructed an agent execution platform on prime of our Kubernetes infrastructure. Each agent runs inside an remoted container that replicates an actual HubSpot developer atmosphere. Brokers compile the code, run automated assessments, learn error outputs, and iterate on their very own till all the pieces works. No human intervention required.

By the tip of Section 2, 96% of engineers have been utilizing AI instruments, engineering velocity was up 60%, and contours of code up to date per engineer had elevated 48%. We have been beginning to ship higher merchandise quicker with brokers. However that was just the start.

Section 3: Scaling with our AI Platform (Mid 2025-Current)

HubSpot’s platform strategy to product growth has at all times been how we’ve created extra buyer worth. Once we constructed reporting and automation on the platform stage, we didn’t simply ship one function; we shipped that functionality throughout each hub concurrently. That’s how innovation compounds.

We utilized that very same logic to our AI infrastructure in Section 3. As an alternative of constructing each agent from scratch, we constructed the shared basis as soon as: how brokers entry knowledge, what actions they will take, how they connect with the remainder of HubSpot. All the pieces runs on prime of it.

The result’s that each one of our brokers are interoperable. They converse the identical language, share the identical toolsets, and draw from the identical context. A buyer will get a constant expertise no matter which agent they’re utilizing as a result of, beneath, they’re all constructed on the identical infrastructure. And since they’re all linked, each new functionality we add makes the entire system extra beneficial. That’s one thing a group of level options can’t replicate.

And it was made potential by how we’ve scaled engineering with AI. Right this moment, 100% of our engineers use AI, strains of code up to date per engineer are up 73%, and time-to-first-feedback on pull requests has dropped by 90%. Meaning much less time ready and extra time transport issues prospects really use.

Why this issues: Compounding buyer worth

Having the correct infrastructure accelerates the tempo of innovation. For HubSpot, each agent we construct makes the platform extra highly effective. Every bit of context we add to the platform makes every agent simpler. For purchasers, which means the product retains getting higher, quicker, and extra linked.

What used to take months now takes weeks, and people weeks translate immediately into new capabilities within the palms of entrepreneurs attempting to succeed in the correct viewers, reps attempting to shut offers, and Buyer Success Managers attempting to retain prospects. They don’t want to consider the platform beneath. They merely get to expertise the outcome.