Glossary
Words we use, in plain terms
Short answer first, then a link if there's a longer story. The entries on the DPDP Act describe how we read it so far and aren't legal advice. Where an entry says we haven't settled something, we mean it.
A2A
A2A, short for Agent2Agent, is an open protocol that lets agents built by different teams talk to each other. One agent asks another to do a piece of work and gets an answer back. Google announced it in 2025. It covers how agents converse. Whether the agent on the other end is who it claims to be is a separate question.
Accountable owner
The person or company that answers when an agent does something. If an agent files a GST return with a wrong figure, the accountable owner is who the tax office would ask. It may or may not be the company that built the software. It is whoever sent the agent out. Every credential we are planning names one.
Agent
On this site, software that acts for someone without being told each step. It books the slot, files the form, applies for the loan. As we read Indian law, an agent is a tool and not a person, so a human or a company has to answer for what it does. Agents often call other tools, and sometimes other agents.
Agent card
A short self-description that an agent publishes, in the style of the A2A protocol: its name, what it can do, where to reach it and how to authenticate with it. Think of a shop signboard with an address. It tells you what the agent says about itself. It doesn't prove any of that, or say who stands behind the agent.
Agent identity
The answers to three questions about a software agent. Which agent is this? Who sent it? Is that still true today? A name alone isn't enough, because anyone can type a name. A hospital deciding whether an agent may reschedule a surgery slot needs to check all three before the request goes through.
AgentFact
AgentFacts is the name the NANDA project uses for a self-published description of an agent and what it can do. We mention it because people ask how we relate to it. As we understand it, it's a peer effort in the same area, not a rival. We haven't compared the two closely yet.
Attestation
A signed statement by someone in a position to know. A bank saying this agent is ours is an attestation. So is a regulator saying this company is licensed. It is worth as much as the person making it and the trust the checker places in them. It can also be wrong, which is why it needs an expiry date and a way to withdraw it.
Audit trail
The record of what happened: which agent did what, when, and on whose say-so. After a bad filing or a wrong payment, it lets people work out the sequence instead of arguing about it. A trail kept only by the company being questioned has an obvious weakness, so we care about who holds a copy.
Consent manager
Under India's DPDP Act, a registered entity through which a person can give, review and withdraw consent to share their personal data, from one place. Think of one window instead of a form at every service. How a consent manager fits when an AI agent acts for you is something we are still working out. Not legal advice.
Credential
A signed statement an agent carries that says something checkable: who it works for, what it may do, until when. A driving licence is the everyday version. The person checking it doesn't have to phone the issuer each time, but has to be able to tell it's genuine and hasn't been cancelled.
Data fiduciary
A term from the DPDP Act: whoever decides why and how personal data is processed, alone or with others. Say you ask an agent to book a doctor's appointment and it handles your health details. The fiduciary could be the company that built the agent, the company running it, or the clinic. We are reading the Act on this. Not legal advice.
Data principal
The DPDP Act's word for the person whom personal data is about. If an agent handles your bank statements, you are the data principal. The Act gives data principals rights, such as asking what is held about them and asking for corrections. For a child, a parent or lawful guardian acts for them. Not legal advice.
Delegation
Handing a task to someone else, with permission to act for you in a limited way. When you let an agent renew your insurance, you have delegated. Done well, the agent holds proof of exactly what you handed over and nothing wider. Done badly, it's a shared password, and the other side can't tell what you meant.
Delegation chain
What you get when a delegate delegates again. You ask an assistant agent to arrange travel, it asks a booking agent, which asks a payments agent. The chain is the trail of who passed which permission to whom. We think each link should carry the same authority as the one before it, or less. Never more.
DID
A decentralised identifier is a kind of identifier, defined by a W3C standard, that doesn't depend on one company's directory to exist or be looked up. It gives an agent or an organisation a name that anyone can check using a published method. We haven't settled how DIDs fit with our own work, and we won't pretend we have.
DPDP Act
The Digital Personal Data Protection Act, 2023, India's law on handling personal data. It names the people involved, data principals and data fiduciaries, asks for consent in plain language, and sets duties and penalties. As far as we can tell it doesn't mention AI agents. How it applies to one acting for you is open. This is not legal advice.
Evidence that holds up later
A record a regulator, a court or an auditor could rely on, which is more than a log line saying something happened. It shows what was checked, when, and what the answer was, and it was kept by the party that did the checking. If a hospital lets an agent in on Tuesday, it should be able to show on Friday why.
Fail closed
When a check can't be completed, the answer is no. If the register is unreachable or a credential can't be read, the request waits or is refused. The opposite, failing open, lets everything through when the check breaks. That is convenient until the one day it matters. Banks and hospitals usually want the first.
Interoperability
Different systems working together without a special deal for each pair. A hospital in Pune should be able to check an agent that a bank in Mumbai registered, using the same method. UPI works across apps and banks for this reason. Without it, every company's directory is a walled garden. Getting there takes shared standards, and people willing to adopt them.
Issuer
Whoever signs a credential. It might be a company vouching for its own agent, an industry body saying a firm is licensed, or an office granting a permission. The checker decides which issuers it trusts, and for what. A credential from an issuer nobody recognises is a signed piece of paper and little else.
MCP
The Model Context Protocol, an open protocol first published by Anthropic in late 2024, that lets an AI application connect to tools and data sources in a standard way. It mostly settles how an agent reaches a tool. It isn't a register of who agents are. We haven't done a full comparison with our own work yet.
Proof of possession
Showing that the agent presenting a credential is the one it was given to, and not someone who copied it. A bearer key, such as a plain API token, works for whoever holds it. Proof of possession adds something the rightful holder can do that a thief with a copy can't, so a stolen credential is useless by itself.
Registry
A public list you can look things up in. A vehicle registration system is the picture we keep using: a number leads to an entry saying who owns the vehicle and whether the papers are current. For agents, the entry says who runs the agent, who stands behind it and whether it is still active.
Relying party
The one who relies on a credential to decide something. A hospital deciding whether an agent may reschedule a surgery slot is a relying party. So is a tax portal, or a loan desk. It is the party on the receiving end that needs to check, and the one who will be asked questions if the check was skipped.
Revocation
Withdrawing a credential before it would have expired. A company finds a fault, or the person supervising an agent leaves, and wants the permission off today. The switch isn't the hard part. The hard part is that every office checking the credential sees it's off, quickly. A permission that can only be withdrawn in theory is a risk.
Royalty-free licence
Permission to use something, here the technology behind a standard, without paying us each time. If a standard is royalty-free, any company can build it into its product without a fee. We intend ours to be. A licence has been drafted but not signed, and we won't say it is in place until it is.
Scope
The limits of what a credential allows. Read the account but not change it. File drafts but not submit them. Spend up to a set amount. A scope that is too wide is the usual way agents end up doing things nobody meant. Narrow scopes are boring and good, but somebody has to enforce them, not only trust the label.
Section 8 company
A company registered under Section 8 of India's Companies Act, 2013 for charitable or public-interest aims. Profits go back into those aims, and no dividend is paid to members. We are planning to set up Saakshya as one, which is why we say in formation. It isn't incorporated yet, and no fee is due from anyone.
Task-scoped
Permission limited to one job, rather than an open licence. A credential that lets an agent file this month's GST return for this one business, and nothing else, is task-scoped. When the task ends, so does the permission. If the agent is fooled or misbehaves, the damage stays inside the box it was given.
Verifiable credential
A credential in the format of a W3C standard, built so that someone can check who issued it and that it hasn't been altered, without calling the issuer each time. It is a general format, and degrees and licences are the usual examples. We are looking at how the same idea could apply to agents. We haven't published a format of our own.
Verification
The check a relying party does before it lets an agent in. Is this credential genuine, is it still valid, does it cover what's being asked, and is the agent presenting it the one it was issued to? It happens each time, not once at sign-up, because the answers can change between Monday and Friday.
Witness (साक्ष्य)
The meaning of our name. In Hindi, साक्ष्य is evidence, or the testimony of someone who saw what happened and can be asked about it later. We want every agent action to have a witness in that sense: a record that doesn't depend on the agent's own word. Every agent needs a witness is our line.