'The AI assistants era is over.' So says Harvey, on the back of last week's launch of more than 500 legal AI agents, with Australia among the first markets to gain access. The release also includes Agent Builder, an early-access tool that lets firms create their own agents on top of the catalogue. Customers across the platform have already built another 25,000 custom agents of their own.
At first glance, the number is impressive. On closer inspection, three worries begin to emerge.
A product that needs 500 workflows is telling on itself
If a tool genuinely worked the way the marketing describes, as a digital colleague who performs delegated legal work, you would not need 500 specialised versions of it. You would need one. The reason the catalogue exists, and the reason Agent Builder exists alongside it, is that scoping a task narrowly is still the main lever for getting a usable output from a language model. Pre-built agents are the scaffolding firms reach for when the underlying intelligence is good but not yet good enough to be trusted on a vague instruction.
That is a fine place to be technologically. The problem is that 500 agents is the opposite of intuitive. A user-friendly tool is one where a lawyer can describe what they want and the system gets out of the way. A catalogue of 500 prebuilt workflows pushes the cognitive work back onto the user: which agent is the right one for this matter, which one is closest, which one needs to be modified. The product surface gets bigger and the lawyer ends up doing translation work between their actual task and the nearest available agent.
Adoption friction goes up, not down
The pitch for 500 agents is breadth. The reality for an innovation partner or a head of legal tech at a law firm is that breadth means more work. Someone has to evaluate which of these agents are reliable, which are accurate enough for the firm's standards, which integrate cleanly with internal precedents, and which are essentially decorative. Multiply that by 500 and across practice areas, and you have months of pilot work before the firm has confidence in any meaningful slice of the platform.
The 25,000 custom agents already on the platform make this worse, not better. That number is being framed as evidence of adoption. However, we think it is also evidence that customers are doing significant configuration work to get the tool to fit their practice. Every one of those custom agents represents lawyer hours spent specifying, refining, and validating a workflow that the firm now depends on but does not own. There are versions of this trade that are good for a law firm and versions that are not, and the difference comes down to how portable the configuration is when the firm wants to leave or when the pricing changes.
A tool that requires this much trialling and configuration to deliver value is relocating the friction back onto the customer, rather than removing friction from adoption.
Accountability gets blurry as work becomes more delegable
Harvey's framing is that agents are 'digital colleagues' who perform 'delegated work.' The Clayton Utz quote in the same announcement is more careful: lawyers still apply human creativity, make judgment calls, verify output, and accept accountability.
The risk is that the marketing framing wins inside firms over time. As 500-agent platforms scale, and as the licence costs need to be justified, internal pressure builds to push more work through the system with lighter-touch review. Relevantly, nothing about the legal profession's ethical obligations changes when the work is done by an agent. Confidentiality, competence, supervision, candour to the court, the Federal Court's GPN-AI practice note: all of it still attaches to the human lawyer. Notably, though, what does change is the practical question of who reviewed what, to what standard, and how thoroughly.
The risk is the subtle normalisation of lighter review on work that used to get heavier review, and not many people paying attention until something goes wrong.
Less is more
None of this is a knock on Harvey's engineering: the platform is known to work across a wide customer base. The point here is just that the announcement is selling a product slightly different from the one it describes, and the real product seems to be one that adds friction, rather than alleviating workflow burdens.
Sometimes one sophisticated tool, adopted simply, is worth more than 500. If that sounds closer to how your firm wants to work, try Habeas at app.habeas.ai.
