AI agent handoffs have a published standard, and where that standard puts the moment of consent matters more to your firm than any of the engineering behind it. DeepJudge released the Agent Handoff Protocol on 13 August 2026 as an open specification for moving a piece of work out of one AI application and into another. Artificial Lawyer reported the same day that Harvey is rolling out support in beta and that Thomson Reuters intends to follow. What travels between the two products is a stated objective, a chosen sample of the conversation, and the files the person at the keyboard picked out.
What a handoff moves
The protocol is narrow by design, and its authors are candid about the limits. It does not copy a model's hidden state, it does not pass credentials, and it does not let one product run another product's tools. What it sends is a package holding the objective, a chronological sample of the messages, the resources the user approved, and two identifiers that keep the task stitched together if the work later travels back the other way. The receiving application authenticates the request and then decides for itself how to do the job.
That restraint is sensible, and it still means client material leaves one supplier and lands at another. A resource in this protocol is a file or a record. In your practice that is a witness statement, a lease, a schedule of loss or a run of client correspondence. The specification states plainly that each application applies its own policies, so what arrives at the second supplier sits under that supplier's terms on retention, training and security rather than under the ones you negotiated with the first.
The consent sits with your fee earner
The specification places the check at the keyboard. The user reviews the destination and the included resources before the transfer, and the sending application has to obtain explicit confirmation of the destination, the context and the resources. In a law firm that user is a solicitor or a paralegal in the middle of a piece of work, and the decision being put to them is a disclosure decision about a client's papers wearing the clothes of a convenience feature. Nothing in the protocol asks whether the receiving supplier appears on your approved list, because the protocol has no way of knowing you keep one.
Two duties meet at that button. The second supplier becomes a processor of your client's personal data from the moment the package lands, so your record of processing, your privacy notice and your due diligence file should name it beforehand rather than afterwards. Paragraph 6.3 of the SRA Code does not soften because a transfer took one click, and the person who clicked will be asked what they understood about where the material went.
What to settle before the button appears
Put two questions to each of your AI suppliers in writing. Ask whether they have implemented the protocol or intend to, and ask which destinations are configured for your account and whether an administrator can restrict them or switch the feature off entirely. A supplier who cannot tell you which destinations are open has answered a third question you did not ask.
Then decide who in your firm approves a transfer and write it into the AI use policy you already have. Handling it the way you handle papers going out to counsel or to an expert gives your staff a rule they recognise and a habit they already keep. Ask for the record as well. The protocol carries a thread identifier that both applications hold, so a supplier can tell you what left, when, and at whose request, provided your contract obliges them to.
None of this argues against the idea. Firms already run one tool for research, another for drafting and a third for review, and moving work between them by copying prompts and re-uploading documents wastes time and loses context. What changes now is that the transfer becomes a single approved action rather than a visible piece of effort, and the invisible version needs a rule around it before your staff meet it in a live matter.
The specification is published free to read on GitHub as the Agent Handoff Protocol, released by DeepJudge under the Apache 2.0 licence, and the adoption by Harvey and Thomson Reuters was reported by Artificial Lawyer on 13 August 2026.
If your AI use policy was written before your tools started passing work to each other, this is the clause we add for firms while it is still a paper exercise: talk it through with us.