Harvey, the legal AI company backed by OpenAI and Sequoia, made two announcements worth reading together this month. Agent Builder is the customer-facing release: a product that lets legal teams build domain-specific agents capable of handling multi-step work. Spectre is the internal-facing system: an autonomous agent platform that increasingly runs Harvey itself — handling engineering and operational work that used to require a human prompt. Together they describe the direction Harvey is betting on: professional services delivered by agents that reason through tasks rather than execute templates.
Agent Builder: Workflows That Think
Agent Builder is a direct evolution of Harvey’s existing Workflow Builder, but the core change is what the agents can do with an instruction. Workflow Builder let users chain together pre-defined steps — extract this, summarize that, format into a memo. Agent Builder lets teams construct agents that reason through a task, adapting to what they find rather than following a fixed script.
The architecture keeps humans in the loop where it counts. Agents surface decisions and flag moments where pre-defined user input would improve the result — a design pattern that matches how senior lawyers actually delegate work to associates. For practice areas where the shape of the task varies (M&A due diligence, complex contract review, regulatory analysis) this is substantially more useful than the scripted workflow that came before.
Teams can embed templates to constrain an agent’s outputs — essentially giving the agent a style guide and format contract alongside its reasoning instructions. Agents can be shared across a firm, which turns them into a codified piece of institutional knowledge: the way this team handles privilege reviews, or the way that team structures a closing memo, becomes a reusable asset.
Spectre: Harvey’s Internal Operating System
The second announcement is more strategically interesting. Harvey has been building Spectre, an internal autonomous agent platform, and it is increasingly running the company.
Spectre takes a request — from Slack, the web app, an automation, or the system itself monitoring incidents and customer feedback — and turns it into a durable run. The run executes inside an isolated sandbox, connects to systems like GitHub, Datadog, and Linear through explicit boundaries, and returns reviewable artifacts: summaries, diffs, branches, pull requests.
The key detail: much of what Spectre does is not triggered by a human prompt. Gabe Pereyra, Harvey’s president and co-founder, has described Spectre as increasingly self-directed — the system notices an incident, reads the bug reports, drafts the fix, opens the PR. Harvey engineers review; Harvey engineers didn’t have to ask.
This is a meaningful boundary crossing. Most “agentic” systems still need a human to start the clock. Spectre is describing a world where the agent watches the workload, decides what needs doing, and does it — with humans moving to a review-only role for an expanding share of the work. See agentic AI architecture patterns for the design tradeoffs this raises.
Why This Is a “Law Firm World Model”
Harvey’s framing is that Spectre is the template for a law firm world model — a system that understands the firm well enough to know what needs to happen next. The mapping from engineering workflows to legal workflows is direct:
- Code repositories → legal matters
- Pull requests → review workflows
- Diffs → document versions with provenance
- Sandbox boundaries → ethical walls and client isolation
If Harvey can run its own engineering operation with an autonomous agent platform, the same platform applied to legal work changes the structural economics of a law firm.
The Structural Threats to Law Firms
Harvey is explicit about what this means for how law firms operate:
- Staffing models built around associates handling rote work become obsolete when AI handles document review and research at scale.
- Pricing structures based on billable hours face pressure when tasks that took days complete in minutes.
- Apprenticeship pipelines that trained junior lawyers through repetitive work need redesign when that work disappears.
These are not speculative. The benchmark data in the 2026 Stanford AI Index shows organizational AI adoption at 88% and student AI use at 80%. The pipeline that currently produces senior lawyers runs through years of the exact work an agent does faster. If firms do not redesign apprenticeship, they will have a mid-career talent cliff in a decade.
What Builders Should Take Away
Harvey’s two-track approach — ship a constrained product (Agent Builder) for customers while running an unconstrained autonomous system (Spectre) internally — is a pattern worth noticing. It shows up elsewhere too: Anthropic exposes Claude Opus 4.7 as a general-purpose API while running Mythos internally for cybersecurity research. The pattern is: use the frontier capability to accelerate your own engineering, then ship a reviewed, bounded version of that capability as product.
For any team building AI-native professional services, three practical lessons:
- Human-in-the-loop is not a constraint, it’s a feature. The agents that will actually get adopted are the ones that surface their decisions cleanly, not the ones that automate the whole task and ask forgiveness.
- Agents as durable artifacts. Harvey’s choice to make agents shareable across a firm — codified institutional knowledge — is the thing that turns a smart demo into a retention mechanism.
- The internal platform compounds. Spectre isn’t just a productivity win for Harvey; it’s the substrate the next product version gets built on. Companies without an internal agent platform will find themselves shipping slower than companies that have one.
Sources: Harvey — Introducing Agent Builder · Harvey — Autonomous Agents Are Legal Next · Harvey — Building Spectre · Artificial Lawyer on Gabe Pereyra