A note from the founder
We built Safeguard's tools with agents and AI as first-class users, not as an afterthought. I believe agents will soon use products like ours more than people do. They will read the mail, book the calls, file the documents and pay the small invoices. If that is where work is going, an agent should not have to borrow a person's inbox or calendar to do its job. It needs its own identity, its own workspace and clear limits, and the people responsible for it need to see everything it did and be able to stop it at any time. That is what we built.
Hritik Kumar Sharma, Founder and CEO, Safeguard
The quiet problem with useful agents
Ask a team that has put an AI agent to real work how the agent sends email, and the answer is usually some version of "it uses Priya's account." Ask how it books meetings, and it is the same story with a different calendar. Ask where it keeps the contract it drafted last Tuesday, and someone goes looking through a shared folder that four other people also write to.
None of this is malicious. The tools we all use were built for people, so the shortest path to an agent that can send a follow up is to hand it a person's credentials and let it act as that person.
The cost shows up later, usually at the worst moment. A customer asks why they received an email with the wrong figure in it. Finance asks who approved a payment to a vendor. An auditor asks for evidence that automated steps in a regulated process were reviewed. A security engineer wants to stop odd activity right now, without locking a senior salesperson out of their own mailbox in the middle of a deal.
In each of those moments, a borrowed account gives the same unhelpful answer: a person did it. Maybe they did. Maybe the agent did, because a person asked, or because it read something in an inbound email. The logs cannot tell you, because as far as every system involved is concerned, there was only ever one actor.
Safeguard's answer is simple to state. Every Safeguard agent gets its own workspace: its own calendar, its own seat in meetings (and meetings of its own that people can join), its own mailbox, its own messaging, its own wallet and its own drive. All of it is tied to the agent's own identity, and all of it sits under the same guardrails: scoped permissions, approvals for anything sensitive, spend limits, a per agent kill switch and an activity timeline you can export for audit. Nothing is on by default.
This post explains why borrowed access fails, how each part of the workspace works in plain language, how the guardrails apply across all of them, what a real piece of work looks like when two agents and a person do it together, and a checklist for teams that want to adopt agents without losing control of them.
Why borrowed access breaks three things at once
When an agent works out of a person's accounts, three separate properties break, and each one matters to a different team.
Attribution
Attribution is the ability to say, after the fact, who did something. In most business systems it is tied to the signed in account, so every action an agent takes through a person's account is attributed to that person. That person now carries responsibility for actions they did not take. A weekly review of "what did the agent do" turns into a hunt through one person's sent folder and calendar history. And in an incident, the first question, "which identity did this," has no honest answer.
The note behind Safeguard's agent identity puts it in one line: as enterprises deploy AI agents against their own systems, "who did this, and was it a human or an agent" stops being a curiosity and becomes an audit requirement.
Approval
Approval is the ability to require a person's yes before a specific action happens. Teams want this for exactly the actions agents speed up most: emailing a customer, accepting a meeting with an outside party, paying an invoice, sharing a file outside the team.
When the agent acts as a person, there is no natural place for that step. The mail system sees a person sending their own mail. The payment system sees a person who may well have authority to pay. The approval ends up inside the agent's own instructions, which makes it only as reliable as the model's willingness to follow them on every run. An approval the agent can skip is not an approval. It has to be enforced by the system the agent acts on.
Stopping
The third property is the one people think about least until they need it. If an agent runs on a person's credentials, stopping it means revoking the person's access too, or hunting down the agent process and hoping no scheduled job, queued email or pending meeting join fires anyway. What teams actually want is one switch per agent that stops everything that agent is doing, in every tool, and leaves the people around it untouched.
The common fix
All three problems share a root cause: the agent has no identity of its own. Give it one, give it tools tied to that identity, and enforce approvals, limits and stops at that level, and all three properties come back. That is the design behind the Safeguard agent workspace.
The foundation: every agent is its own principal
Everything in the workspace sits on Safeguard agent identity, and every capability below inherits its rules from it.
In Safeguard, each agent is registered as its own principal in your tenant. A principal is an identity the platform recognises and can hold responsible: people are principals, and so are agents. An agent principal has:
- Its own identity and handle, so it appears by name in logs, approvals, messages and meetings rather than hiding behind a person.
- An owner, the person accountable for what the agent is allowed to do.
- Roles, permissions and tools. Each agent carries its own roles, an explicit list of allowed permissions and an explicit list of allowed tools. A tool that is not on the list cannot be called.
- A ceiling. An agent never exceeds its owner's authority. If the owner cannot do something, the agent cannot either, whatever roles it holds. That closes the loophole of using an agent to do what its owner is not permitted to do.
- A status: active, suspended or deactivated, checked on every action.
- Short lived credentials. Agents act with their own tokens, separate from any person's login, which expire after fifteen minutes. A token carries the agent's identity and the identity it acts for, so every action records both.
Safeguard tracks which agent holds what access, at what level. That is the core of agent identity, and it lets every capability answer three questions: who did this, who approved it, and how do we stop it.
Tenant isolation is built into identity
Your tenant is the boundary around your organisation's data in Safeguard. Every workspace capability takes the tenant from the verified credential of the caller, never from a header or a request field that an agent, or anyone else, could change. If anyone outside your tenant asks for one of your agents, messages, files or approvals by identifier, the answer is the same as if it did not exist: they cannot list it, read it, change it or confirm it is there. Every capability is tested this way: a second tenant tries to read the first tenant's items by identifier and must get "not found" every time.
What each agent gets
With identity in place, each agent gets six tools of its own. Here they are in plain language, one at a time. The guardrails that apply across all of them come after.
Calendar
Each agent holds its own schedule. It accepts invites sent to it and books time inside the hours its owner allows.
Time the agent commits to is the agent's commitment, recorded against its identity. When it books a call with a customer, the invite comes from the agent. Nobody's personal calendar fills up with entries they did not make, and nobody wonders whether a slot was booked by them or by software acting in their name.
The calendar is also where an agent's meetings start. When an agent is invited to a call, or books one of its own, the calendar is how it knows to be there.
Meetings
Safeguard agents join calls as named participants. Each agent appears with its own video tile and its own voice, so everyone on the call can see that an agent is present and which one it is. Up to three agents can take part in a single call.
In the meeting, the agent:
- Keeps a live transcript for the length of the call.
- Answers out loud when addressed by name. This is configurable, so a team that wants a quiet note taker can have one.
- Answers typed questions asked through Ask, for people who would rather not interrupt.
- Keeps a live board of decisions, action items and key points, and can present it to the call.
- Shares its screen when it has something to show, such as the board at the end.
The meeting agent works in Google Meet, including joining automatically from Google Calendar, and in Safeguard Meet, Safeguard's own meeting rooms for signed in Safeguard users. Calls in Safeguard Meet use in transit encryption (DTLS SRTP).
Customers choose the model the agent uses from Safeguard's model list. Transcription providers include a self hosted Whisper option that runs on Safeguard's own servers, with no third party involved. When an agent joins a call, it runs on Safeguard's servers. For people who would rather keep audio on their own machine, Safeguard Desktop offers optional on device transcription and translation; in that mode audio never leaves the laptop, and nothing is sent anywhere unless the user chooses to send the transcript.
Meetings an agent hosts, that people join
Meetings work in both directions. An agent can join a call that a person set up, and an agent can host or hold a meeting of its own that people join alongside it, in Safeguard Meet and in Google Meet.
That changes who does the organising. An agent that owns a piece of work, a customer follow up for example, can book the meeting from its own calendar, invite the people who need to be there and the other agents who can help, and run the call. The people who join get everything the meeting agent offers: the live transcript, Ask for typed questions and the running board of decisions and action items. Nobody needs to lend the agent their meeting link.
Because the agent hosts under its own identity, the meeting itself is attributed to the agent. The invite, the attendance and the notes that come out of the call all trace back to the agent that held it and to the owner responsible for that agent.
Each agent has its own mailbox for sending and receiving. It is not a forwarding rule on a person's inbox or a shared alias. It belongs to the agent.
- Approval rules for anything sensitive. Outgoing mail that meets your approval rules is held as pending until a person approves it. The agent can draft, but it cannot send that message on its own, and the hold is enforced by the mail system, not by the agent's prompt.
- A full audit trail. Every message sent and received, and every action the agent takes on a thread, is recorded against its identity.
- Attachment screening. Inbound attachments are checked against an attachment policy before they are stored. Blocked attachments are never passed to the agent's other tools.
A mailbox can also be paused on its own, so an agent can stop sending mail while it keeps working elsewhere.
Email is the channel where an agent's mistakes reach people outside your company most directly. When the mail comes from the agent's own address, the recipient knows who they are dealing with, your team knows which messages to review, and replies go to a mailbox the agent and its supervisors can see.
Messaging
Each agent can send and receive messages. That covers three kinds of conversation.
Chat with people. People in your team message an agent the way they message a colleague, and the agent answers. Files sent in that conversation, or produced by the agent in it, are filed in the agent's drive with a note of the conversation they came from. Chat channels your team already uses, such as Slack, can be connected to an agent so that mentions and direct messages reach it and its replies post back.
Messages between agents. Agents talk to each other. One agent can hand a piece of work to another, ask it a question or pass it a document. Every message is tied to the identity of the agent that sent it and the agent that received it, is recorded in both agents' activity timelines, and is covered by the same approval and kill switch rules as everything else. A suspended agent cannot send or receive. An agent cannot use another agent to reach a tool or an action it is not allowed itself, because each agent acts inside its own permissions and its own owner's ceiling.
Text messages on the agent's own number. An agent can have its own phone number. Text messages to that number land as messages for the agent, and calls to it are routed to the agent's voice. Each number belongs to one tenant. Opt out replies such as STOP are honoured, and the existing do not contact and consent records apply to every send.
The reason to give messaging to the agent's identity, rather than routing it through a person's chat account or phone, is the same as for mail. When a customer texts the agent's number, or a teammate messages the agent, or one agent asks another for help, the conversation belongs to the agents involved and shows up in their records.
Wallet
Each agent can have its own wallet. The agent pays within limits its owner sets, and any payment above those limits needs a person's approval before it goes through.
The wallet is built on a double entry ledger, the standard accounting approach where every movement of money is recorded as a matching debit and credit, so the balance is always explainable. Your team funds the wallet by topping it up. The agent cannot add money to its own wallet, and it cannot spend past its limits without a person saying yes. Wallet activity has its own audit record, and finance can download a statement of every movement as a spreadsheet file.
This answers the question finance raises the moment agents start buying things: how do we let it pay without giving it the company card? A separate pot of money, with its own limits and record, belonging to one agent and frozen in one step.
Drive
Each agent has its own file space inside Safeguard Assets for the documents it creates and receives. It uses the same file store, sharing model and version history Safeguard Assets already provides, with each file marked as owned by the agent.
Files reach an agent's drive from:
- Documents the agent writes, such as a summary or a draft proposal.
- Meeting notes. When a meeting the agent attended ends, the summary, action items and transcript are filed as one document.
- Meeting recordings, only when a separate recording setting is on. It is off by default and depends on meeting bots and meeting consent both being enabled.
- Mail attachments in and out. An inbound attachment that passes the attachment policy is copied with a note of the message it came from; a blocked one never is. A copy of each attachment the agent sends is filed too.
- Chat files sent to the agent or produced by it in a conversation.
Every file records where it came from, so a document traces back to the meeting, message or conversation that produced it.
Access is narrow. The agent reads, writes, versions and deletes only its own files, and its identity comes from its credential, never from the request, so one agent cannot claim to be another. The owner can edit and change sharing. A tenant administrator can view, delete and set quotas. Everyone else sees an agent's file only if it is shared with them, and the default is private. Every write, sharing change and deletion is recorded.
Each agent has a storage quota set by the tenant administrator, with a per agent override; a write that would exceed it is refused and leaves nothing behind. Deleted files are held for thirty days, then removed for good. Files past your tenant's retention period are removed by a nightly job, bytes first, then the record.
The guardrails that apply to all of it
Six tools with their own controls would still be six places to look. The important design choice is that the guardrails sit across the whole workspace, at the level of the agent's identity, and work the same way everywhere.
Nothing is on by default
Every workspace capability has its own switch, and every switch starts off for every tenant. A capability is turned on deliberately, and you decide which agents use it through their allowed tools and permissions.
Off really means off. When a capability is off for your tenant, its routes answer as if they do not exist, its tools are not offered to your agents, its menu entries and tabs are hidden, and no request is made from the page. If the system that holds those switches cannot be reached, the capability behaves as off. This is failing closed: an outage or a misconfiguration leaves you with less agent capability, never more.
There is one deliberate exception. Suspending an agent, pausing a mailbox or stopping a campaign never sits behind a capability switch. Turning a capability off can never take away your ability to stop an agent.
One approval inbox
Approvals already live in each tool: held mail, wallet payments over the limit, agent tasks with sensitive tools, changes to an agent's own access, social posts waiting to publish. The approval inbox brings them into one view. Each item shows the tool, the agent, a short summary of what it wants to do and when it asked. A person approves or rejects from the inbox, one at a time or in bulk, and the tool that owns the action carries out the decision.
Two details matter. The decision stays with the tool that owns the action, so the inbox is one place to look, not a second copy that could drift. And if a tool cannot be reached when you open the inbox, that source shows as unavailable, never as empty. "Nothing to approve" and "we could not check" look different, because they are different.
Spend limits
Each wallet has limits set by its owner. Payments inside them go ahead and are recorded; payments over them are held and appear in the approval inbox with the amount and payee. The wallet enforces the limits, not the agent. An agent cannot raise its own limit or fund itself, and a payment split into pieces shows every piece in the ledger and the timeline.
The per agent kill switch
Every agent can be suspended or deactivated in one action, and administrators can change several agents at once. Suspension revokes every credential the agent holds, and it also stops what the agent already has in motion:
- Its mailboxes are paused, so queued mail does not go out.
- Its active outbound activity, including texts and calls, is stopped.
- Its scheduled meeting joins are cancelled.
- Its wallet refuses further spending.
Each stop is safe to repeat, and any action the agent tries afterwards is refused. Because the switch acts on the agent's identity, it never touches the people around it, and it leaves every other agent running.
The activity timeline and audit export
Every action an agent takes is recorded against its identity: access events, mail, messages to people and to other agents, payments, tasks, meetings joined and hosted, outbound activity. The activity timeline merges them into one view per agent, in time order, with the agent shown as the actor on every line. The agent task record is hash chained: each entry carries a fingerprint of the one before it, so a gap or an edit in the middle is detectable.
Administrators can export an agent's timeline for any date range as a spreadsheet file or as structured data. The export is itself recorded, so pulling evidence leaves evidence. This is the file your auditor asks for when they want to see what an automated process did and who approved it.
Privacy and erasure
Agents handle personal data: names in mail, voices in meetings, numbers in texts. The workspace ties into Safeguard's privacy paths, so a confirmed erasure request reaches agent data too. When a meeting participant's data is erased, the agent's notes and recordings from that meeting are removed. Mail origin files in an agent's drive are covered by the same erasure path as email keyed data elsewhere. Retention follows the media and text settings your tenant already uses.
Worked example: two agents and a person run a customer follow up
Guardrails are easier to judge when you watch them work. Here is one realistic piece of work done by two agents and one person, with every checkpoint shown.
The cast:
- Maya, an account executive.
- Griffin, an agent Maya owns, which handles her customer conversations.
- Sentinel, an operations agent owned by the finance lead, which handles vendor invoices and filing.
1. The customer asks for a follow up
After a first call, the customer's security lead emails Griffin asking for a follow up with their engineers next week. The email lands in Griffin's own mailbox.
Griffin checks its calendar, finds a slot inside Maya's allowed hours and drafts a reply proposing Tuesday at 10:00 in Safeguard Meet. Because it is going outside the company, it matches Maya's approval rule and is held. The approval inbox shows: "Griffin wants to reply to the customer proposing Tuesday at 10:00." Maya approves. The mail goes out from Griffin's address, the customer accepts, and the meeting sits on Griffin's calendar with Griffin as host.
Griffin also messages Sentinel: "Customer follow up Tuesday. They may need the proposal translated. Can you check vendor availability and cost?" The message carries both identities and appears in both timelines.
2. Griffin hosts the call, the people join
On Tuesday, Griffin opens the meeting it booked. Maya joins, the customer's security lead and two engineers join, and Griffin invites Sentinel as well, because pricing for the translation may come up. Two agents and four people are on the call, each agent with its own tile and voice.
Griffin keeps the live transcript. An engineer asks, "Griffin, what did we agree last time about where the data is stored?" and Griffin answers out loud from its notes of the earlier call. Maya types a question into Ask rather than interrupting. When the translation comes up, Maya asks, "Sentinel, what did the vendor quote?" and Sentinel answers from the reply it received that morning.
At the end, Griffin shares its screen and walks everyone through the board: three decisions, five action items with owners, two open questions. The notes, with the summary, action items and transcript, are filed in Griffin's drive, marked as coming from that meeting.
3. Handing work between agents
Two action items belong to Sentinel: order the translation and file the signed documents. Griffin does not do them itself. It messages Sentinel with the two items and passes the meeting notes as a document. Sentinel acknowledges and starts.
That handover is a recorded event in both timelines: which agent asked, which agent accepted, what was passed. If something goes wrong later, nobody has to reconstruct who was supposed to do what.
4. The follow up mail
Griffin drafts the customer follow up: a short recap, the decisions, the action items with owners and dates, and the notes attached. It is held for approval. Maya corrects one date in the draft and approves. The mail goes from Griffin's mailbox, and a copy of the attachment is filed in Griffin's drive as sent with that message.
The customer's engineer replies by text to Griffin's number with a quick question about the start date. The text lands as a message for Griffin, Griffin answers from the agreed plan, and the exchange is on Griffin's timeline next to the email thread.
5. Paying the vendor within limits
The translation vendor emails its invoice to Sentinel. The attachment passes the attachment policy and is filed in Sentinel's drive as an inbound document. The amount is inside the limit the finance lead set, so Sentinel pays from its wallet, and the payment is a line in its ledger and its timeline.
Had the invoice been over the limit, the payment would have been held and shown in the approval inbox with the amount and vendor for the finance lead. Sentinel could not raise its own limit, top up its own wallet, or ask Griffin to pay on its behalf from a different wallet without that request appearing as a message in both timelines, and Griffin's own limits and approvals would still apply. At month end, finance downloads Sentinel's wallet statement and reconciles it like any other account.
6. Filing the documents
When the translated proposal arrives, Sentinel files it in its drive, marked with the message it came from, and passes a copy to Griffin. Griffin sends it to the customer, held for Maya's approval like any outside mail. Sentinel writes a short deal record document listing what was ordered, paid and sent, and the finance lead shares it with the team. Everything else stays private to the agents and their owners until someone chooses to share it.
7. When something goes wrong
Now the uncomfortable version. Suppose an inbound email to Sentinel carried hidden instructions trying to get it to send documents to an outside address, and Sentinel started drafting a message to Griffin asking it to forward the files.
A security engineer sees the odd activity and suspends Sentinel. Every credential Sentinel holds is revoked. Its mailbox is paused, its pending message to Griffin never sends, and its wallet refuses further spending. Griffin keeps working, because the switch acts on one identity. Maya and the finance lead keep their own accounts and carry on.
The engineer exports Sentinel's timeline for the past week: every message received and sent, to people and to Griffin, every approval and who gave it, every payment and every file. The investigation starts from a complete record. Attribution, approval and stopping all held, because each agent had its own identity and its own tools.
What this gives security, compliance and finance teams
Security teams get a clean boundary per agent: its own identity, fifteen minute credentials, an explicit tool list, a ceiling at its owner's authority, and a switch that stops it everywhere without touching people or other agents. Capabilities fail closed. Agent identity pairs with Safeguard's MCP Guardrail and MCP security SDK, which adds under 10ms of network latency, for teams that also want guardrails on their agents' tool calls.
Compliance teams get evidence. Most controls come down to who has access to what, who approved this, and can you show me the record. With agents as their own principals, those answers exist for automated work in the same form they exist for people. Safeguard itself is SOC 2 Type I and Type II audited.
Finance teams get a contained way to let agents pay: a separate ledger per agent with its own limits, approvals and statement, frozen in one step.
The people who work alongside agents get their accounts back. Their inbox holds their mail, their calendar holds their meetings, and nobody has to ask them whether they sent that email.
A checklist for adopting agents safely
Whether you use Safeguard or not, settle these before an agent touches real mail, calendars, messages, money or customer documents. Each item is a yes or no question.
Identity
- Does every agent have its own identity, separate from any person's account?
- Does every agent have a named owner who is accountable for it?
- Is it impossible for an agent to do more than its owner is allowed to do?
- Do agent credentials expire quickly, and are they separate from any person's login?
- Can you list, for every agent, which tools and permissions it holds?
Defaults
- Is every agent capability off until someone deliberately turns it on?
- If the system holding those switches is unavailable, does the capability turn off rather than on?
- Has each agent been given only the capabilities its job needs?
Approvals
- Have you written down which agent actions need a person's yes? Consider external email, meetings with outside parties, payments above a set amount, sharing files outside the team and any change to an agent's own access.
- Are those approvals enforced by the system the agent acts on, not by its instructions?
- Is there one place where a reviewer sees every pending approval across every tool?
- Does an unreachable source show as "unavailable" rather than "nothing pending"?
Messaging between agents
- Is every message between agents tied to both identities and recorded in both timelines?
- Can an agent reach a tool or action through another agent that it could not reach itself? The answer should be no.
- Does suspending an agent stop its messages to other agents as well?
Money
- Does each paying agent have its own wallet, separate from company cards?
- Are limits set by a person and enforced by the wallet, with the agent unable to fund itself or raise them?
- Can finance download a statement and reconcile it?
Mail, texts and meetings
- Does each agent send mail and texts from its own address and number?
- Are opt outs honoured on every text, and are inbound attachments screened?
- Does an agent appear under its own name in every meeting it joins or hosts?
- Have you chosen the agent's model and where transcription runs?
Files and data
- Do agent files live in a space owned by the agent, private by default, with a quota?
- Can you trace each file to the meeting, message or conversation it came from?
- Do retention and erasure reach agent files, notes and recordings?
- Is recording a separate, deliberate choice with consent in place?
Stopping
- Can you suspend any agent in one action, revoking its credentials and stopping its queued mail, texts, meetings and payments?
- Does the kill switch work even when a capability is off?
- Does stopping one agent leave people and other agents untouched?
- Have you run a drill before going live?
Evidence and isolation
- Can you see one agent's full history across every tool, in time order, and export it for a date range?
- Is the export itself recorded, and would your auditor accept it?
- Is every action scoped to your organisation from the verified credential, and has your vendor shown one customer cannot read another's agent data by identifier?
A team that can answer yes to every item can let its agents do real work in real systems and still explain, approve and stop everything they do.
Common questions
Can people join a meeting an agent set up? Yes. An agent can host a meeting in Safeguard Meet or Google Meet, and people join alongside it with the live transcript, Ask and the board of decisions and action items.
Can agents see each other's files? Not by default. An agent reads and writes only its own files. It gets another agent's document when that agent passes it in a message, and the handover is recorded in both timelines.
What happens to an agent's work when it is suspended? Its credentials are revoked and its queued and scheduled actions stop. Its records stay: mail, messages, files, ledger and timeline remain for review.
Does the agent record every meeting? No. Recording to the drive is a separate setting, off by default, that depends on meeting consent being enabled. The transcript and board are part of the meeting agent itself.
Which model does the agent use? You choose from Safeguard's model list.
Where this is going
Agents already do work that used to belong to a person: reading the inbox, setting up the call, writing the recap, chasing the invoice, filing the paperwork, and increasingly, asking each other for help. The question is no longer whether agents will do this work. It is whether they do it under their own name, inside limits someone set on purpose, with a record that holds up when someone asks what happened.
That is what an agent workspace is for. Each Safeguard agent has its own calendar, its own seat in meetings and meetings of its own, its own mailbox, its own messaging, its own wallet and its own drive. Each is tied to the agent's identity, scoped by its owner, held for approval where it matters, capped where money is involved, stoppable in one action and recorded in a timeline you can export. Nothing is on until you turn it on.
To see how agent identity and the agent workspace fit your team, talk to us at https://safeguard.sh.