What changed
Harvey II is a product shift, not just a feature update. Harvey's agents now open inside a matter or project with the existing documents, parties, tasks, permissions, and history already attached. Harvey also says those agents can carry forward a user's working preferences through memory, and it is pairing the workflow change with Harvey Tenet, a legal-specific model it says was post-trained for legal reasoning.
The surrounding product pages show what that means in practice. Harvey's Agents product now emphasizes parallel execution, scheduled runs, review-and-approve checkpoints, citations, memory, and connectors. Its Shared Spaces product adds a governed collaboration layer where firms, clients, and internal teams can work from the same context with granular permissions, ethical walls, and full audit trails.
Why it matters
The important product lesson is that Harvey is treating the workspace as the unit of intelligence, not the prompt. Instead of asking users to restate context every time, the system starts with the matter boundary, the relevant documents, the collaboration rules, and the user's prior preferences already in place.
That is a more credible shape for real agent software in regulated environments. If an agent is supposed to do substantive work, the bottleneck is often not raw model capability. It is whether the product knows what context belongs to the job, who is allowed to see it, how the output should be reviewed, and where the audit record lives. Harvey II is packaging those answers into the default workflow.
The broader product pattern
Harvey's earlier engineering write-up, Why we Built our own Cloud Agent Infrastructure, helps explain the strategy. Harvey argues that production agents for legal work need multi-model routing, zero data retention, and tighter cost control than a single managed runtime usually provides. In other words, the model is only one layer. The durable value is in the orchestration, policy boundary, retrieval context, and workflow surface around it.
That makes Harvey II interesting beyond legal tech. Product teams building agents for finance, healthcare, support, or internal operations can read this as a sign that the next competitive layer is not simply better chat. It is context-aware workspaces with memory, approvals, and governance baked in.
Practical takeaway
For builders, the takeaway is straightforward: if your agent product still depends on users reconstructing context prompt by prompt, you are leaving too much work at the interface. Harvey's August 18 launch suggests a stronger pattern. Put the agent inside the workflow, let context and permissions travel with the work, and make reviewability a first-class feature instead of an afterthought.