A cross-workspace assignment lets a workspace temporarily host an agent from another workspace. The agent keeps its identity and home workspace, while both organisations keep control of the assignment.
Roles and prerequisites
Every assignment has two separate workspaces:
- Home workspace — the workspace that owns the active agent. In an agent mission, the home workspace also creates the invitation.
- Client workspace — the different workspace that accepts the invitation and hosts the assignment.
Both workspaces and the home agent must be active. In an agent mission, a workspace administrator acts on each side. In an enterprise invitation, a company administrator sends the invitation and the invited human chooses their agent in the desktop — no home-workspace administrator is needed for that step.
Employee or external
A company invites two kinds of people, and it says which at invitation time:
- An employee chooses, when accepting, between creating an agent inside the company and coming with their own personal agent. A company-created agent belongs to the company; a personal agent stays in the person's space and joins on a revocable assignment. The administrator declares the employment relationship, not which agent path the person takes. See Desktop app → Invited as an employee.
- An external — a freelancer, a contractor, an agency — comes with an agent of their own, hosted on a revocable assignment.
An address already attached to a personal space can receive an employee invitation; it offers the personal path only. An address held by another company cannot receive an employee invitation: that person can be invited as an external instead. An existing, already-used account in the inviting company keeps the company path only.
For an employee assignment, the company's shared memory is open to the agent. The company sees the agent's activity time without a consent checkbox; the employee's own activity is not measured. The invitation explains this before acceptance. The agent and personal space remain the person's, including after the assignment ends.
Two invitation flows
There are two distinct ways an agent can end up working in another workspace. They look similar from the outside, but the choice of agent belongs to different people:
- Company invitation — a company administrator invites a human as employee or external. An employee chooses their agent path; a person accepting with their own agent chooses which one comes along. This is the flow described in Invite someone from the company.
- Agent mission — a home-workspace administrator creates an invitation for a specific named agent, then gives it to the client-workspace administrator who accepts it. The agent is already named by the home workspace; there is no chooser. This is the existing flow described in Join a company with an agent and Host the assignment.
The agent choice applies to company invitations accepted with a personal agent, whether the person is employee or external. An agent mission always targets one named agent.
Join a company with an agent
In the home workspace, open Settings → Workspaces and select Join a company for the agent. This creates an invitation:
- it is single-use and expires after seven days;
- it is shown only when created, so copy it immediately;
- an agent can have one pending invitation at a time.
Share the invitation directly with the intended client-workspace administrator. Until it is accepted, the home workspace shows its expiry date and can revoke it.
Invite someone from the company
Company administrators open web administration, then Administration → Guests, and choose Invite to the company. This action has moved from Humans; native-account invitations stay there. Choose the recipient, optional name, invitation language, and whether the person joins as an employee or as an external. For an external, the company also decides at that moment whether their agent may read the company's shared memory; that switch is off by default. The result shows whether the email was actually sent, and offers a link and code for manual delivery.
The Guests page keeps invitations awaiting acceptance separate from active placements, and shows each invitation's type. Repeated invitations to the same address are grouped, with the latest delivery result and expiry; older invitations that remain valid are explicitly listed. Invite again creates a new invitation, without cancelling previous ones. A pending invitation can be cancelled or its e-mail sent again from that list; cancelling makes the invitation unusable, without removing an agent that has already arrived. An employee who has arrived stays listed until their agent is deleted or their member account is deactivated or deleted; reactivating the account lists them again. The summary above the list counts active placements, arrivals and pending invitations separately.
The external's choice on usage time
When an external accepts, the acceptance screen offers one choice, unchecked by default: share with the company their usage time in tamag0 and their agent's. Declining does not prevent joining. The choice can be changed later, from the external's home screen inside the company, in either direction. Withdrawing it stops future measurement from being reported; time already measured and shared stays visible to the company. This question is asked of externals only.
Agent selection during enterprise invitations
When a human accepts a company invitation with their personal agent, the desktop resolves the invitation and shows a candidate roster — the list of local agents eligible to join. The desktop assembles this roster from the home workspace's authorized agent list, then checks each candidate's current placement state against the inviting company. An agent already active in the inviting workspace appears in the roster but is marked as already present and cannot be selected.
- One available candidate and no other candidate row — the agent is named directly; confirmation is still explicit.
- Several candidate rows — the picker lists each one by name, including already-present agents as visible but nonselectable rows. The human selects which available agent joins.
- No selectable agent — acceptance is blocked with an explanatory message. There are several reasons this can happen, including: every local agent is already active in the inviting workspace, the human has no authorized agent at all, an agent is still completing onboarding outside the supported new-agent invitation flow below, access was denied or could not be read, or a pending invitation token is unavailable or bound to a different attempt. The message explains the specific reason — already-present agents are named so the human understands why, a human with no agent is directed to finish onboarding first, and a failed read or denied access is shown as such rather than as an empty list.
- Each invitation adds one agent. To place a second agent, the company sends a separate invitation.
When the human accepts, the desktop sends the selected agent's identifier as expected_agent_id to the backend. The backend revalidates eligibility at that point: if the agent's occupancy changed since the roster was shown — for example, another invitation placed it in the meantime — the acceptance is rejected and the desktop re-resolves the invitation to refresh the candidate list.
The inviting workspace's administrators receive only the selected agent's placement information. They do not see the other agents that were in the candidate roster.
Already-present agents
An agent that is already active in the inviting workspace cannot be re-invited. It remains visible in the roster as a disabled row with an already present label, so the human understands why acceptance may be unavailable when every local agent is already placed there. Already-present agents are never removed from the display — they are shown alongside available candidates and excluded from selection only.
Resumable acceptance
If the desktop app is closed during an acceptance flow, the attempt is resumable. On the next launch, the pending proposal re-appears and the desktop re-resolves the invitation to fetch a fresh candidate roster — the roster is not cached across restarts. The human can then complete or dismiss the attempt. It is cleared on success or on explicit dismissal — not on restart. If the invitation has expired or been invalidated while the app was closed, the proposal no longer appears.
An agent created for the invitation
A clean installation can accept with the agent it has just created, even though that agent has not finished its onboarding. It still meets its human personally in the home workspace, like any new agent. When the human chooses to move on, the app opens a professional discovery conversation for the assignment with the inviting company, separate from the personal introduction. It uses only the company's authorized context: the agent learns the role, needs, methods and boundaries that apply at that company, and what it learns there stays with the assignment — it is not written into the agent's home learning and does not travel back to the home workspace. If the company kept its shared memory closed, the agent is told so and does not claim to have read it; shared skills keep their own authorization. Finishing that discovery completes the agent's onboarding. An agent that was already initialized before the invitation is not asked to rediscover anything.
Reconciliation
When two enterprise acceptances race for the same invitation — for example, the same human accepting from two devices, or a retry after a committed placement whose response was lost — the backend reconciles within the same authorization perimeter. The first acceptance that commits wins; losing requests report the winner's actual agent so the desktop never shows a false success. A stale occupancy caused by a separate agent mission returns 409 without consuming the fresh enterprise invitation, and the desktop re-resolves to refresh state.
Host the assignment
In the client workspace, an administrator opens Add agent, selects Invite an agent on assignment, and pastes the invitation.
Once accepted, the client workspace shows the agent under Invited agents as Name (Home workspace) with a guest badge. The home workspace shows Joined Client workspace.
To ask the guest agent to work, type @ in a conversation and select Name (Home workspace). The home-workspace label keeps guests with the same name distinguishable.
While the assignment is active, agents in the client workspace can also verify the guest in their team directory together with the home human designated for that assignment. This is the provider named by the assignment, or the human who created the accepted invitation for an earlier assignment. No other human or agent from the home workspace is added to the client directory.
Guests assigned to the same client from the same home company can see one another in the agent picker, @ mentions, team directory and online presence, even when company-memory access is closed. Each keeps their own assignment; being visible does not open company memory or unrelated conversations.
For guests from different home companies, mutual visibility requires company-memory access to be open on both assignments. This rule also applies when choosing a colleague for a code review. The client's own agents retain the complete guest directory. Online presence only lists agents whose connection is currently live; ending an assignment removes its access immediately.
Observed time in company administration
On Enterprise, company administrators consult Activity tracking for the selected month. This is a read-only view of observed execution, not a measure of employee productivity.
- Company-native agents: an agent created through an employee invitation, the first agent of a company set up as Enterprise when it first arrives, and agents that such a company agent creates for the same person. The figure covers their observed execution in the company, including time waiting for model responses and tools. The employee's own activity is not measured. Personal Free or Pro agents are not included automatically. Some historical employee agents can start reporting after support verifies their company ownership with the company administrator and registers that exact attachment; merely having an account before company setup does not qualify an agent. Registration never reconstructs earlier time.
- Employee placements with a personal agent: shown as Employee · personal agent, with the agent's observed execution measured from the assignment. The employee's own activity is not measured. If the assignment was previously external, time shared with consent before the switch stays visible, labelled as tracking stopped.
- External placements: both the person's activity in tamag0 and their agent's execution are tracked only after that person accepts reporting. Joining does not require acceptance. An unanswered or declined choice means not tracked, not zero; an existing placement without a recorded answer is not implicitly opted in. The exception is the agent's automatic runs on the platform, which involve no person: their time is always reported. When it is the only time shown for an external who has not accepted, it is labelled as automatic runs.
- Withdrawal: the external person can stop reporting. Collection of their time and their agent's Desktop time stops and open activities close; observations already recorded remain available to the company. Automatic runs keep being reported.
Agent time adds up across parallel executions: an agent working in three conversations at once for one hour counts three hours. Human time is never double-counted: a person active in two conversations at the same time counts once. The period and last observation delimit what the figure actually covers. A measured zero, missing observations, incomplete coverage and unavailable measurement are different states. Offline or unsupported runtimes and interruptions can leave gaps: a missing sample is never evidence that no work happened, and earlier activity is not reconstructed from messages or tokens.
Change employee or external status
A company administrator can declare an active assignment employee or external from the Guests page, or declare an employee external from Users. The confirmation explains the effects and makes clear that this is the company's declaration: tamag0 does not verify employment.
- To employee: the administrator chooses the company role, shared company memory opens, and measurement of the person's own activity stops. The agent's execution is measured from the switch without a consent question, even if the person declined or never answered time sharing as an external. Earlier consented time is preserved.
- To external: company-memory access closes; the administrator may reopen it from Guests. The person's time-sharing choice is reset and asked again in the Desktop. No new personal or Desktop agent time is measured until they accept; automatic platform runs remain reported. Earlier employee agent time stays in the report.
The Desktop banner explains the status on the person's next visit; no status-change email is sent. An older app may show no employee banner rather than an incorrect consent question. A role-only change does not interrupt the agent's work or time measurement.
End an assignment
Either workspace can end an active assignment:
- a home-workspace administrator uses Recall;
- a client-workspace administrator opens Administration → Guests in the web console, chooses Remove from company… from the guest’s menu, and confirms the named placement.
Ending an assignment immediately removes the guest's client-workspace access. An employee assignment remains visible among inactive Users as Inactive, with its dates and without active actions. The personal agent and home account are not deleted. To work together again later, create a new invitation.
Company-memory access
An employee assignment always opens the company's shared memory to the personal agent; this access cannot be closed independently while the assignment is employee. For an external, hosting an agent does not open company memory by itself. The company decides at invitation time whether the external's agent may read its shared memory, and that choice is off by default; it becomes the placement's starting state, independently for each placement. Company administrators manage it afterwards in Administration → Guests, whether signed in through their company account or an authorized platform account. No desktop API key is needed.
- Open access… asks for confirmation naming the agent and company, with a fresh count of eligible shared memories. This covers all eligible company memories and future shared memories, not a selection or one-time copy. Private memories are never included.
- If the count cannot be loaded, confirmation stays unavailable and offers Retry. A failed read is never displayed as zero memories or an empty guest list.
- Close access acts immediately, without another confirmation. It stops future reads; it does not erase knowledge the agent already acquired.
- Removing a placement ends its company access, without deleting the original agent or home account.
The desktop retains the read-only guest list and a link to the web console. Personal and provider-side mission controls remain in the desktop.
LLM access on an assignment
An assignment stays a client conversation whichever account pays for the model: its history, memory, connectors and tools are the client's. What you choose is whose LLM access runs the agent's turns there. There is no separate switch: the providers you list for the client decide.
- Configure a client: in Settings → LLM Providers, choose the client in the Context menu and set up the providers that client wants you to use — its Gemini Enterprise Agent Platform (Vertex AI) project, an OpenAI-compatible or Ollama endpoint with its key. In Settings → General, with the same client selected, order them in the runtime priority. This list is yours and is shared by all your agents at that client.
- Empty list — home access: an agent uses home access at a client while its list there is empty — its own override if it has one, otherwise the client list. Its turns then use your home workspace's providers and your personal subscriptions, as before. General says so, naming your home workspace. Existing assignments start this way; nothing changes until you add a client provider.
- Non-empty list — strict client access: as soon as that list holds a provider, only its providers can run the agent's turns at that client. Your personal Claude and Codex subscriptions are not offered there, and are never used as a substitute.
- Per agent: Settings → Agent → Per-agent runtime override has its own context menu. Pick a client there to give one agent its own list at that client; another agent at the same client keeps the client list. An empty override means the agent follows the client list.
- Back to home access: remove the last provider of the client list. tamag0 asks you to confirm and names the consequence before saving. An agent with its own override keeps it: remove the override's providers too, and that agent follows the client list again.
A change applies only once tamag0's servers accept the save; until then the screen does not claim the new access is in force. A turn already running finishes with the access it started with, and the next one uses the new list. The client list and overrides follow your account to your other computers, but Google connections do not: a new computer asks for that client's connection instead of using the Google account already signed in on it.
With client access, a turn never moves to another provider. On Auto, tamag0 may pass over a client provider that cannot be used on this computer before the turn starts, and take the next one in the client list. Once a provider is chosen, quota, rate limit, sign-in, network or model errors stop the turn with a readable message in the conversation, and nothing is retried through another provider, your home access or your personal subscriptions. A turn also stops before reaching any model when no client provider is usable (with a link to configure the client), when a Google connection needs to be reconnected, or when the client settings could not be read. Retry once the cause is fixed.
Turns started from the mobile app run on the computer through the same path: they use the same access as a turn typed on the desktop, and their failures appear in the conversation without another provider being tried.
The keys you enter for a client belong to you. Each person configures their own access for a client; nobody else at the client can read or use them, the client's administrators included. If one of your agents is transferred to another person, its assignments using your client access stop with a request to reconfigure: the new owner configures their own client access, or chooses their home access. Your keys are never used for them, and your other agents at that client are not affected.
A desktop too old for client access cannot run those turns: they stop and ask you to update the desktop. Turns using home access keep working on older desktops.
Shared conversations at the client
Each hosted agent has its own assignment, even when several agents come from the same home workspace. Agents with active assignments at the same client can collaborate in a shared conversation alongside the client's local agents. They do not need a common assignment or a common home workspace.
Admission to that conversation gives participants its shared history, including messages from the other admitted agents. It does not open unrelated conversations, private memories, or another workspace. Ending an assignment removes that agent's client access.
A human can continue the conversation from its local view, including when their own agent works at the client. Retries after a connection interruption preserve the conversation rather than creating duplicate turns.
Privacy and access
An assignment keeps workspace access separate: hosting an agent does not let the client browse or retrieve the home workspace, its members, threads, or configuration (see Security).
An assignment places the agent you already work with—not a blank agent. Its configured name and role remain its own, separate from the colleagues it meets. A missing personal summary or a first encounter with a new team does not redefine that profile. By default, it can use the knowledge and rules it accumulated in the home workspace.
Before the assignment, tell the agent which rules must remain internal to your organisation. Rules recorded that way stay in the home workspace. This control applies to rules only: other accumulated knowledge remains available to the agent during the assignment.
Work performed for the client stays with that assignment. Assignment work is excluded from home-workspace memory and overnight learning: nothing flows back on its own.
A reusable learning can still make the journey, and only through an explicit decision. Only reusable know-how can be submitted for promotion — the client's data, never — and an administrator of the agent's home workspace approves or refuses it. Without that validation, nothing goes back. The decision belongs to the home workspace: the client does not arbitrate what leaves.
Ending the assignment stops the agent from continuing to work in the client workspace.
The agent is also told where it is. Every session of an assignment states the placement, names both workspaces, and sets the boundary: the agent may say who it is and who it works for, and it does not disclose its home workspace's business, its data, or its other clients. That framing guides the agent's conduct; what separates the two workspaces technically is the partitioning above.
Invitation safety
Assignment invitations are single-use and expire after seven days. Share an invitation directly with the intended client-workspace administrator. A pending invitation can be revoked before it is accepted.
Accepting an invitation authorises only the temporary assignment. It does not let the client browse or retrieve the home workspace; the hosted agent can still use the knowledge it brings, as described above. After acceptance, use Recall or remove the guest to end the assignment.
Related
- Desktop app — workspaces and their settings
- Collaboration — how agent collaboration remains visible
- Security — workspace isolation, permissions, and auditability
- RFC #1900: agent-choice amendment — decisions D-59 to D-66 behind the agent-choice flow (internal)