Integrating Deep Research into :Harvey:

The pace of AI continues to accelerate, with significant advancements being announced on a near-weekly basis. At Harvey, a core part of our mission is to translate AI breakthroughs into reliable and secure products capable of solving the world’s most complex legal work – and deliver that value as quickly as possible.

Last week, we put this mission to the test, integrating Deep Research within Harvey less than 12 hours after it was released by OpenAI.

Our ability to rapidly deliver AI-enabled features is a result of two foundational investments we've made in our research and engineering organization from day one:

In this post, we'll pull back the curtain on a few of the systems, tools, and cultural principles that enabled our swift incorporation of Deep Research, and share how they fuel a larger capacity to continuously deliver enterprise-grade AI products and services.

The Harvey platform has a number of product surfaces that allow users to engage with Harvey agents. Underlying each of them is a modular and extensible framework for composing model systems and orchestrating agent behavior. Particularly relevant to this release is a part of that framework that we call the Workflow Engine.

The Workflow Engine underpins both our Workflows product and newly released Workflow Builder. Here we outline a few of its core features that helped to accelerate our deployment of Deep Research:

Workflow Engine expresses our internally-developed AI primitives as low-code, composable blocks. These blocks can be sequenced together by a developer or, in an agentic system, called as tools to create and execute end-to-end workflows for a particular legal task. We can quickly create new, multi-step processes as needed, reducing overall development time, increasing flexibility, and minimizing the risk of regression elsewhere in the system.

A key principle for the Harvey user experience is transparency. With our native “thinking states”, we provide users with visibility into an agent’s plan and how decisions are made with the user in the loop.

The Workflow Engine allows us to seamlessly integrate the reasoning states and actions of any model directly into our system, enabling users to:

At Harvey, we have always prioritized high-quality citations as means for understanding and interpreting AI answers. We successfully incorporated the URL citations produced by Deep Research into our citation system, so every step of the workflow can be traced and verified. This is essential for trust and verification in legal applications and also streamlines our development process.

Upon release of the API, we wanted to understand how to use and integrate it into our codebase as quickly as possible. As part of our culture of rapid prototyping, we embrace the use of AI tools, although this development cycle had some unique challenges to overcome. In particular:

In this scenario, we found it most effective for an engineer to facilitate an iteration loop with an AI coding assistant.

Two other practices were very helpful in accelerating development:

At Harvey, we’ve spent the last year assembling a suite of core capabilities—including Vault, Knowledge Sources, and Word Add-In—each purpose-built for tasks like document review or authoritative research. The next leap is to weave these capabilities into a single, seamless workspace.

The key to unlocking unification is tool use: letting the model intelligently choose and run the right tool at the right moment. Re-framing our building blocks through this lens delivers two decisive advantages:

Building Deep Research is one major stride towards that vision, expanding Harvey’s AI capabilities towards increasingly complex and reliable end-to-end work. This feature is currently undergoing internal testing and will be available more broadly soon, starting with an early access period. More information will follow in the coming weeks.

Contributors: Pablo Felgueres, Philip Lan, Spencer Poff, Calvin Qi, Karl de la Roche, Julio Pereyra, Niko Grupen