Interoperability is standardising. Retention is not.

Say a small clinic chain in Pune gives a scheduling agent to its front desk. It reads appointment requests, checks doctors' calendars and sends reminders. A hosted runtime does the heavy lifting, so the chain never built servers of its own.

Eight months later, a patient writes in. She'd like to know where her messages went, and she'd like them deleted.

The front desk asks the IT lead. The IT lead asks the vendor. Somewhere in the second step it turns out that nobody wrote down which country the conversation logs sit in, or for how many days.

That's the shape of a problem that isn't getting much attention, because the attention is elsewhere.

The part that is being standardised

Over the last year the industry has been busy agreeing how agents talk. MCP covers how an agent reaches tools and data. A2A covers how agents talk to each other. In August the Agentic AI Foundation announced (17 August 2026) that A2A had become one of its hosted projects, alongside MCP. The foundation is hosted by the Linux Foundation, and its post says A2A is backed by over 150 organisations. We've seen press figures for the foundation's total membership that are higher, but we couldn't open them, so we're not quoting one.

That's real progress, and good for anyone who doesn't want to be locked in. We'll call it the wire problem: how do two pieces of software understand each other?

The foundation's own site also lists an Identity & Trust working group, which says it covers portable identity, delegation and how permissions flow between agents. So the wire isn't the whole of what's being worked on. It is the part with settled projects.

The part that isn't

The retention problem is different. Where does the data go once the agent has handled it, and how long does it stay?

No protocol answers that. It's decided privately, runtime by runtime. As a current example, OpenAI's Agents API overview, which we read on 12 October 2026, says the API supports data residency only in the United States and doesn't support Zero Data Retention. It adds that picking a self-hosted sandbox doesn't change that. The API is in beta, so this may change.

OpenAI's data controls page lists the same endpoint with a 30-day retention for abuse monitoring, application state kept until deleted, and no use of the data for training. It doesn't say how a customer deletes that application state, and we'd want that spelled out.

Nothing in that is sinister. It's a vendor describing its product clearly, which is what we'd want. Our point is that the next vendor will describe something different, and a buyer has to go and read each page.

What India's law says about it

We're reading the Digital Personal Data Protection Act 2023 here, from reproductions on dpdpa.com because the official PDFs wouldn't open for us. This is not legal advice. Please check the Gazette text.

Section 16 takes a blacklist approach. The Central Government may, by notification, restrict a fiduciary from transferring personal data to a notified country. Absent such a notification, a transfer isn't barred by this section, and other Indian laws with stricter rules keep applying. So sending data to a US-only runtime may be permitted. That isn't the same as being off the hook. Section 8(1) says the fiduciary is responsible for processing done by it or on its behalf by a processor, whatever the contract says.

On deletion, section 8(7) requires erasure when consent is withdrawn or the purpose is no longer being served, whichever comes first, unless a law requires retention. It also requires the fiduciary to cause its processor to erase data it handed over. Section 12(3) adds an erasure right a person can exercise by request. Taxmann reports that these sections come into force on 13 May 2027.

For the clinic, that means the patient's request is the clinic's to honour, even though the logs sit in somebody else's system.

Four questions before you adopt a runtime

Where is conversation and tool-call data stored, and can you see the country from the contract or the console?

Can retention be turned off, or can you set your own deletion window? Ask for it in days. "Short" isn't an answer.

Can you reach into the vendor's system to carry out an erasure request, or do you have to ask them and wait? Ask how long they take.

And does the contract make the vendor's data handling your legal problem or theirs? Often it's both, and you want to know in what proportion before a patient writes in.

None of these is hard to ask. They're easy to forget because the demo is about capability and the pricing page is about price.

What we don't know

We don't know how many buyers ask these questions today. We haven't surveyed anyone, and we won't pretend otherwise.

We don't know whether retention will ever be standardised the way the wire has been. Regulators may push it. Vendors may compete on it. It may stay a contract detail forever.

We don't know how the Data Protection Board will read section 8(7) for data sitting in a vendor's hosted state. And the OpenAI position above is a snapshot from one day. We'd correct it if they change it.

One question to leave with you. When you evaluated your agent runtime, did retention and data location come up, or only capability and price?