Everyone has an agent now

Two things happened in the last two weeks of August 2026.

On the 18th, Harvey shipped Harvey II, built around memory — the system learning an individual lawyer's drafting style and word choices and carrying them across Word, Outlook and its agents, plus matter-centred Spaces holding documents, tasks and work history.

On the 25th, Google launched Gemini Enterprise for Legal in preview, with legal-specific skills for brief drafting, citation verification, regulatory scanning, discovery and DSAR fulfilment. Launch customers named are Cleary, Freshfields, Weil, and Williams & Connolly.

We had just finished counting agents across the market when both landed. Here is what the field actually looks like.

Fifty products, four different meanings

In our catalogue of 264 products, 50 are agents or agent infrastructure. That makes it the largest single thing happening in legal AI right now. It also makes "agent" close to meaningless as a product description, because it covers at least four distinct things.

A chat assistant, renamed. Luminance's Lumi summarises contracts and answers questions about them. Useful. Not autonomous in any sense the word usually implies.

A multi-step workflow runner. Harvey Workflows, Legora Workflows, Patlytics Agent, Clerq Agents. You define a sequence, it executes over documents. This is the bulk of the category, and it is closer to scripting than to delegation.

A builder. LegalFly's Agent Studio, Newcode's Aurora, DeepJudge's AI Workflows, Bryter Workflows, Josef, Checkbox AI, Zuva Create. Here the product is not an agent at all — it is a factory for making them, sold to teams who will define the work themselves.

Something actually autonomous. Flank runs agents that handle recurring legal requests inside Outlook, Teams, Jira and Salesforce without a person opening an app. Streamline's Velo Copilot runs "interactively or ambiently." Hona's Lia chases clients and escalates only when it needs to. Eve's communication agents place outbound phone calls for case status and medical records.

The autonomous ones are doing boring work, and that is the point

Notice what the genuinely autonomous agents are doing. Not arguing motions. They chase people who have not replied, answer the same routine question for the fortieth time, request records, place calls, and log what happened.

That is the correct place to start. It is high-volume, low-judgment, verifiable work where a mistake is embarrassing rather than malpractice. The agents doing substantive legal work are still, almost without exception, running under a lawyer who reviews the output before it goes anywhere.

Anyone selling you an agent that practises law unsupervised is selling you a liability.

The sleeper development is a protocol, not a product

The thing we did not expect to find: legal data is becoming callable from outside legal software.

Trellis ships a Claude MCP connector and a ChatGPT plugin for US state court records. midpage ships an MCP connector so Claude, ChatGPT and Perplexity can query its case-law corpus. Twin1 runs an MCP server exposing a lawyer's own context to other tools. Clerq offers API and MCP access to patent search. GC AI sells API access with no seat attached.

The implication is worth sitting with. If the research corpus, the court data and the firm's own knowledge are all reachable by a general-purpose assistant, then the interface a lawyer works in may not belong to a legal software vendor at all. The legal company becomes the data and the domain logic; somebody else owns the window.

Which is roughly what Google just announced.

Google's position is the interesting part

Read the Gemini Enterprise integration list carefully: Docusign, Everlaw, iManage, NetDocuments, RelativityOne, Thomson Reuters — and Harvey, and Legora. Google did not launch a Harvey competitor. It launched a layer that sits above Harvey and Legora and treats them as sources.

For a firm already paying for two of those, that is attractive. For the vendors, being an integration inside somebody else's enterprise agent is a materially different business from being the place the lawyer opens every morning.

One name on that list is worth stopping on, given what we found about tools for self-represented litigants: Courtroom5, which builds for people without lawyers, sits in it alongside Thomson Reuters, iManage and Relativity. It has raised about $420,000, mostly in grants. It was also a launch Justice Partner in Anthropic's Claude for the Legal Industry in May.

That is the connector layer doing something the funding market has not: a company with almost no capital reaching the same distribution as companies with billions, because what it takes to be there is a working MCP server and a corpus worth querying.

Who watches the agents

If agents proliferate, something has to check them. Across all 264 products we found exactly one built for that: Norm Ai's Supervisory AI, a verification layer that monitors whether other companies' AI agents are behaving lawfully.

One product. Against fifty agents and counting. That gap will not stay open long.

The number nobody publishes

We went looking for agent reliability benchmarks — task completion rates, error rates, how often a human has to intervene — and found none. Not from Harvey, not from Legora, not from Google, not from anyone.

We published our own accuracy numbers with the caveats because a claim without a corpus and a method is marketing. Agents are heading into that same territory, one step further from verifiable: at least a transcript can be checked against a recording.

Until somebody publishes how often these things finish the job correctly, "agentic" is a description of architecture, not of results.

Product counts are from our own catalogue, accurate as of late August 2026. Nothing here is legal advice, and we have not tested these agents.