Two Types of Legal Work. One Agentic Platform.

Agents have become the dominant frame for legal AI this year, and it's worth grounding the conversation in what that actually means in practice.

Harvey has been running legal work agentically for a while: 700,000+ tasks are executed daily, over 50 million terms are extracted weekly. That experience has shaped a clear view of what it takes to do agentic legal work well, and where the hard problems actually are. What's changed recently isn't the concept, it's what agentic execution is now capable of and how that changes what's possible in legal work.

Two Kinds of Legal Work. Two Kinds of Execution.

Not all legal work is the same, and the way an agentic platform handles it shouldn't be either. Legal work falls into two categories. There's work that's repeatable, where the goal is consistency and scale. And there's work that isn't: complex, open-ended matters that require sustained reasoning, and judgment across many steps. Those two categories have always existed. What's changed is that both can now be handled agentically, and they each demand a different kind of execution.

For repeatable work, agents follow structured steps for defined tasks, repeatedly and reliably. Contract review checklists. Regulatory monitoring. Compliance checks. Term extraction across hundreds of documents simultaneously. These agents encode a firm's or legal team's best practices into repeatable execution. They run the same way every time, at scale, without variance.

For complex, long-horizon work, agents operate differently: planning, reasoning, and executing across many steps. Analyzing a full contract suite across a transaction. Drafting memos that synthesize multiple knowledge sources. Running end-to-end diligence extractions. Navigating the lifecycle of a fund formation.

These agents work with the lawyer, not around them. The agent generates a plan. The lawyer reviews and edits that plan before execution begins. As the agent works through the prompt, lower-stakes decisions get logged and are available for review. Higher-stakes decisions, anything outside the playbook, anything that materially changes the direction of the work, cause the agent to pause and ask. The result is a transparent record of work done, decisions made, and judgment calls flagged for human review.

These two types aren't a progression. Firms and in-house teams need both running in parallel, because the work demands both. One mode handles the volume: recurring, high-frequency work that has to be done consistently every time. The other handles the complexity: work that requires sustained reasoning across many steps. A capable legal AI platform needs to run both.

Why Now

Foundation models can now sustain complex reasoning over longer tasks, maintain coherence across many steps, and produce higher-quality outputs with less variance than was possible even a year ago. This changes what agents can deliver end-to-end, and in the case of legal.

Harvey has been building toward this kind of execution for a while. Deep Analysis was an early version of this. What's changed is that the quality bar Deep Analysis demonstrated is now achievable across a much broader set of tasks. The infrastructure was already in place. The models have now made this reliable enough to deploy broadly.

Where This Goes

Legal work spans a wide range, from high-volume, structured tasks to complex, multi-step reasoning problems. Agents built for only one of those categories will eventually hit their limits.

The more durable direction is a platform that runs both well: agents that encode and scale the team's best practices alongside agents that handle the work that requires genuine reasoning. With the infrastructure to give agents the right context, the governance to make it enterprise-safe, and the compounding knowledge that makes the platform smarter the longer you use it.

Agents that do the work. Lawyers that provide direction and judgment. That’s what Harvey is building. Learn more about Harvey Agents.