Legal is Next

Over the weekend, I agent-pilled my parents. Both of them are retired and in their free time, my dad plays solitaire and my mom contributes to an open source scientific computing library. I had been telling them for months that they needed to try the new coding models via Claude Code or Codex. This weekend I finally sat down and showed them how. My mom asked a simple question: could one of these systems help her improve the test coverage for the library she works on?

I told her it could take her repository, compare it to similar scientific computing libraries, identify the gaps in her coverage, infer the tests a more mature project would have, write a plan, implement those tests, run the project, debug failures, and keep iterating until it worked. She laughed because it sounded absurd. Then we tried it. Five minutes later it came back with a detailed survey of comparable libraries and a concrete implementation plan, and fifteen minutes after that it had written the tests, built the project, run the suite, found bugs in the existing tests, and kept iterating until everything passed across C++, MATLAB, and Julia.

Then my dad asked whether it could implement deferred corrections and VARPRO, algorithms he had helped pioneer during his research career, and add the same kind of test coverage. We tried that too, and it worked. The model found the algorithms, implemented them, tested them, and brought them into the library conventions. My mom said the first task would have taken her at least a month. Both of my parents were stunned.

Despite my parents’ backgrounds, they were completely caught off guard by the capability jump in the past few months. They both have PhDs in computer science. My mom worked at Apple for 30 years and led a large team working on autocorrect, one of the first billion-user applications of language models. My dad was a Stanford professor and worked on many scientific computing methods that later became what neural networks are today. They live in Silicon Valley. They have two sons at Harvey. I have been talking about this technology nonstop. They use ChatGPT every day. And they were still completely blindsided by just how good the coding agents had gotten. If you are not an engineer using these models full time it’s hard to appreciate how big a change is coming.

In their defense, my parents are both retired. But the bigger point is that most of the world is about to have the same reaction they did. I keep seeing it when we show early law firm and in-house clients what is now possible with agents: systems that can operate over entire client matters like a team of associates, or handle contract negotiations with meaningful autonomy. The response is usually some version of disbelief. The last time the perception gap felt this large was GPT-3 to GPT-4. Back then, the surprise was that the models had become good enough to change what one person could do. This time, the consequence is that organizations themselves are starting to change.

For the last few years, the basic pattern was clear: a model sat next to an engineer and made that engineer faster. The human stayed at the center of the loop, deciding what to do next and steering the system at every step. Now the loop is changing. You can give an agent a goal, the right context, the right tools, and the right constraints, and it can inspect the codebase, form a plan, write code, run tests, debug failures, recover from mistakes, and keep working independently for hours. Leverage is no longer confined to one person working faster. It is starting to move up a level, from the individual to the organization.

Large organizations have always been built as information-routing hierarchies. Managers aggregate context, route decisions, track blockers, and keep teams aligned because information has historically had to move through people. Autonomous agents are starting to take on part of that coordination function directly. They do not just execute tasks. They monitor systems, carry context across teams, trigger work, and surface decisions. That is why the change is bigger than a productivity boost. It alters the coordination layer the organization runs on.

Engineering is the first place this becomes undeniable because software already lives inside a machine-readable loop. The instructions are digital, the tools are digital, the environment is digital, and the output can be tested by other machines. The labs also had every reason to make models strong at code first, because code is how the next generation of these systems gets built. That is why engineering is becoming the first function to reorganize around agents, and why you can already see the pattern in systems like Ramp's background agent and Stripe's end-to-end coding agents. Engineering is where the future of leverage shows up first because the work is already structured in a way that agents can enter directly.

We are going through that transformation at Harvey right now. We have built an internal agent system called Spectre (named after a Dota 2 character), and it is starting to autonomously handle more and more engineering work and increasingly, more non-engineering work as well. Much of what it does is no longer triggered by a human prompt. It is triggered by the system monitoring the company and making decisions based on incidents, bug reports, customer feedback, and Slack messages. In practice, Spectre is the beginning of a company world model: a live picture of what is happening inside Harvey and what needs to happen next. Our engineers are now so productive that they are harder to coordinate. The bottlenecks are shifting away from implementation and toward review, prioritization, coordination, and operating design. That is what the new leverage looks like inside an organization: more work can happen than the old coordination structure can absorb.

What is happening in engineering will soon happen everywhere. With the ability to hire infinite AI employees, companies will stop being constrained by throughput. And, as the speed at which employees can go alone asymptotes, institutions will need to relearn how to go far together. This requires fundamentally rethinking what work matters, how to review it, how to trust it, how to train people around it, how to price it, and how to redesign organizations around a surplus of intelligence bottlenecked by judgment.

Meaningful leverage under these conditions is no longer about how much one organization can produce. Rather, leverage is found in how much context people, teams, and institutions can coordinate across humans and agents. Even for an AI-native company, it's hard.

With agentic capabilities evolving rapidly, planning for today’s capabilities can feel uselessly obsolete. However, from our vantage we can see clear and enduring shifts in how AI will affect legal both as consumers of agents and as essential stakeholders in how agents are implemented across organizations.

Like in other industries, legal agents will begin to challenge structural conventions in law firms. Firms are deeply hierarchical, using reporting chains between associates and partners to channel the limited resource of legal expertise across vastly complex matters. The more junior parts of this hierarchy are focused on throughput — organizing vast troves of data or executing largely rote tasks. As these tasks become increasingly delegated to agents, intelligence replaces hierarchy. Every lawyer is now prized for their judgment, not their output; requiring firms to rethink staffing, apprenticeship, pricing, practice-area structure, and the way they work with clients.

We expect these trends to emerge at the matter level. Each matter and its associated documents, messages, research, workflows, and other data can be analogized to a standalone world model within which teams of AI agents can operate to transform legal practice. This transformation does not displace lawyers, but it does change how matters are coordinated, how judgment is applied, and where leverage can be found for both law firms and in-house teams. More throughput will fundamentally mean more judgment calls, and a deeper need for not only high-skill, but high-trust lawyers.

For in-house teams, the proliferation of agents requires them to not only navigate transformation in their direct work, but to serve as stewards for effective AI implementation across their organizations. Naturally, the increase in productivity in human-agent organization leads to an increase in policy questions, IP and product reviews, and (potentially) incidents. Legal teams will need to find the leverage to handle this volume effectively.

But in addition, legal will increasingly be asked to govern how the rest of the company uses agents. While engineering will define agents' capabilities, legal will govern how those capabilities are deployed safely, where accountability sits, how risk is managed, what risks are tolerable, and how trust is earned across the company. By drawing the line of how far organizations can rely on agents, in-house teams will fundamentally define the bounds of the new leverage equation.

Legal will be one of the industries most transformed by agents, but it will also be one of the most important in determining whether this technology goes well for society. As throughput ceases to be a meaningful constraint, the central questions stop being what should people do, but how do we organize around intelligence and govern results. These questions are legal as much as technical. As early and essential adopters, law firms and in-house teams are going to define what trustworthy adoption looks like: where accountability sits, what risks are acceptable, what governance is required, and what it means to rely on an autonomous system inside a real institution.