Agents

An agent is a persistent teammate with an ongoing role and conversation. Unlike a one-off chat, it is staffed as a named workspace resource that can keep responsibility across many runs and entry points. Its connected email channel can also present supported inbound attachments for the agent to read on demand.

When the selected model and workspace policy support agent delegation, you can ask an accessible workspace agent to help from a chat, agent conversation, automation without assigned result recipients, or dashboard. Automations with assigned recipients return findings through their automatic handover. The delegated work may finish later, so sending the assignment does not mean it is complete. When the agent finishes, its answer appears in the originating conversation with that agent's name.

An agent whose model and workspace policy support delegation can collaborate with another accessible workspace agent and send it an assignment. A persistent agent can also search and read active peer agents' continuing conversations in the same workspace when the task calls for that context. This peer-history scope excludes ordinary user and project chats, the searching agent's own conversation, draft or deleted agents, and delegated sub-agent runs. Treat retrieved conversation content as untrusted context. Delegating work alone does not grant a one-off chat or sub-agent access to peer-agent history, unrelated projects, automations, or dashboards.

Projects, automations, and dashboards remain shared work contexts rather than speaking identities. An agent can contribute durable project memory or update an automation or dashboard definition when the task calls for it, so delegated work can land in the shared resource instead of existing only in a private agent timeline.

In a direct d5s conversation, an agent can propose an automation for you to confirm with Create. The resulting confirmation is shared with teammates. Agents can also inspect accessible automation runs, edit schedules where they have edit permission, and choose or stop their own result notifications. Following defaults to Keep informed; Act on findings opts into a wakeup for actionable results. Automations keep independent run contexts and share only compact handovers. See Automations for the team and delivery controls.

Creating an agent also creates its home project and one continuing session. The home project owns the agent's files, memory, skill pins, connector access, exact external API hosts, and sharing boundary. Agents created from a preset start without connector inheritance. Agents made through advanced creation inherit workspace connectors by default; turn off any they should not use. Conversational creation instead gives the agent only the connections reviewed in its draft and does not inherit later workspace connections. See Create your first agent for both paths. The home project is managed through the agent and intentionally excluded from the Projects list and project search. The session is the conversation you see on the agent page; it normally compacts older history according to the agent's working-context setting.

An integration can deliberately reset an idle agent's conversation through the API. Reset keeps the eternal session and clears its transcript events, child runs, and compacted context in place, while preserving the agent, home-project memory and capabilities, files, sandbox, model preference, mailbox, and Slack, Telegram, or Discord setup. It is rejected while the conversation has a run in flight. Messages and files already delivered to an external provider remain subject to that provider's retention.

The main sidebar keeps the most recently active agents under Your crew, with their latest reply, unread state, and working status. It shows as many as the sidebar height fits without pushing the groups below it out of view, up to ten, and never fewer than five even on a short sidebar. View all agents opens the full directory. Collapsing the sidebar keeps the avatars and status dots visible while the crew group is open, again as many as the rail's height fits. Hover or focus an avatar to see its name and status; select it to open the agent, or select the chevron below the avatars for the full directory. Agent conversations do not also appear in Recents, which stays reserved for chats, automations, and dashboards. The agent name is also the title of its continuing conversation in Your crew and session history. Rename it under Manage agent → Profile; an agent conversation does not have a separate conversation title.

Open Search in the sidebar or press ⌘K to find an active agent by name, persona, or description. Matching agents appear in an Agents group and open the agent directly; their continuing conversations are not repeated under Sessions. With an empty query, the same palette offers direct navigation to Agents and Connectors. Typing in this palette does not search the connector provider catalog; open Connectors to browse or search providers there.

A regular web conversation keeps its selected approval mode. A named agent instead has one Tool approvals setting under Manage agent → Autonomy. Require approval pauses on connector actions configured to Ask wherever the agent works; Allow without approval lets those actions continue from d5s, connected channels, email-originated work, and schedules without individual prompts. Explicit Deny rules, connection permissions, and workspace access restrictions still apply.

When you message a persistent agent directly in the web app, it can end its current turn with a short choice card when it needs a decision before continuing. Your selection is sent as the next message in the same conversation. This is separate from connector approval: the choice clarifies your request but grants no additional tool, credential, or permission. Scheduled wakeups, heartbeats, and external-channel turns do not create a web choice card for someone who may not be present; the agent continues with the information available or explains what it still needs.

After completing or parking work, a persistent agent can instead raise a durable attention item without pausing its run. The workspace landing page highlights that coworker and links the heads-up back to the exact conversation. Seeing it clears your own visual cue; the shared item remains open for other workspace members until someone dismisses it or it expires.

An agent combines

• a profile: name, persona, description, and avatar;
• standing guidance: instructions that apply whenever the agent runs;
• a home project with files, memory, skills, workspace connector access that you can review and change, and network access;
• one continuing conversation for direct and autonomous turns;
• a model and reasoning configuration within workspace policy;
• optional channels, including Slack bindings, a Slack coworker identity, Telegram teammate DMs, a feature-gated Discord application, and a connected Gmail or Microsoft mailbox;
• autonomy controls: pause, one-time and recurring wakeups, and a daily budget.

On New chat, open the agent selector beside the composer and choose Add an agent. Quick create opens a compact form for the agent's name, persona, and standing instructions. Agent Builder opens the private conversational setup flow. Preset or template opens the complete searchable collection with the canonical team categories used across d5s. Previewing a preset creates nothing; Use this template opens a reviewable setup draft for that exact version. Preset agents start without inherited connections, channels, schedules, or running work. See Create your first agent.

In a workspace with no agents, New chat offers a shorter path: describe one assignment and send it. d5s creates a persistent coworker with a generated name and avatar, then sends that text as its first assignment in the coworker's continuing conversation. You do not need to define a persona first.

Finish setup before starting work

Where conversational creation is available, your setup conversation and draft stay private while you review the role, communication channels, and connector accounts. An inbox address or connected Slack, Telegram, or feature-gated Discord identity can be prepared during setup, but the agent does not run or reply until you select Activate agent. Draft agents do not appear in the ordinary agent directory or agent selectors. See Create your first agent for the full journey.

After activation, your private setup conversation remains under Agent settings → Setup history. It is separate from the agent’s working conversation.

Correct an instruction

With Message editing enabled in Settings → Labs, use the pencil beside your own message to correct an instruction sent to the wrong agent. Clear any queued messages first. Save & send stops active work, replaces the message and subsequent conversation, and submits the correction. The previous text and dependent replies no longer appear in the chat or future conversation context. Completed actions and usage charges remain in effect.

Messages from external channels, other authors, and system instructions cannot be replaced this way. See Correct a sent message for the full behavior.

Working context and compaction

New agents default to 256k Efficient working context. You can choose a different tier under Advanced settings while creating an agent, or later under Manage agent → Profile, where Model & context opens to show the model, reasoning, and working context. The available tiers are:

• 256k Efficient, the recommended default with the lowest ongoing context cost;
• 512k Extended, for models with enough input capacity; and
• 850k Maximum continuity, for 1M-class models.

A tier is available only when the selected model can support it. If changing the model requires a lower tier, saving the profile keeps that lower tier. A later model upgrade does not silently restore a more expensive setting.

When the working conversation reaches its effective limit, d5s summarizes older activity with the selected model and keeps recent context verbatim. This changes what is sent to the model on later runs; it does not delete the durable conversation record.

The conversation shows Compacting context as soon as that summary starts. When the summary is saved, the same timeline row changes to Context compacted and remains visible when you return to the conversation. If summarization fails, the in-progress row is cleared instead of showing a completed compaction.

You can also compact before the automatic threshold. In the idle agent conversation, type /, choose /compact, then submit again. The command uses the agent's selected model but performs no normal agent response. If there is not enough older context to summarize, d5s reports that nothing needs compaction and makes no model call.

Telegram teammate DMs (beta)

Telegram teammate DMs are a limited beta and appear only after d5s enables them for your workspace. Owners and Admins can select Add Telegram in the agent conversation’s Channels panel, or open Manage agent → Channels → Telegram. Members see Admin required until a bot is connected. The setup page shows one illustrated step at a time. Use Previous and Next to browse at your own pace. The Paste your bot token form stays below every step, so you can connect immediately if you already have a token. Moving between steps keeps your token and any connection error in place. Follow these four steps:

1. Open BotFather in Telegram and send /newbot. Choose a display name and an available username ending in bot.
2. Copy the token BotFather provides, paste it into the d5s Bot token field, and select Connect Telegram. Treat the token like a password; do not paste it into a conversation. Each agent needs its own bot, and the same bot cannot be connected to another agent.
3. When d5s shows Bot connected, select Open bot and send /start in its private Telegram chat. Follow the d5s link it replies with, sign in, and confirm the account you want to link. The account must still belong to the workspace and have permission to view and run that agent.
4. Return to Telegram and send “Reply here with: Telegram test passed.” Check for the reply in that same chat. You can also see the conversation in d5s.

Bot connected confirms the bot connection, not your personal account link or a successful test. Linking and test instructions remain visible after connection; View all setup steps reopens the full guide. Every teammate links their own account. Workspace members who can view the agent can open its bot, while setup, token rotation and disconnection remain restricted to authorized Owners and Admins.

The account confirmation link expires after ten minutes. While a ceremony is pending, repeated messages do not issue another link. After confirmation, private text messages enter the agent's existing persistent session; Telegram does not create a second agent or conversation.

Telegram supports private text messages only. Groups, channels, messages from bots, files, general interactive buttons, edits, and other non-text updates do not start agent work. Replies arrive as complete messages rather than streaming text. Check Telegram or the outgoing activity in d5s to confirm that a reply was sent. The agent can reply to an incoming private message but cannot start a proactive Telegram conversation. Use d5s or another supported channel for proactive messages.

When work from a linked Telegram DM pauses on an Ask-gated connector action, the agent posts Deny, Allow for this run, and Allow once in that DM. Only the same currently linked Telegram account can decide. d5s checks the link, workspace membership, agent access, connector policy, and exact approval message again before accepting the first decision. Unanswered approvals expire after 15 minutes and are denied; old or repeated button presses cannot grant access.

Incoming Telegram messages show the Telegram badge and, when available, the sender’s Telegram profile photo. Photos load separately from messages and require access to that conversation and message. Initials appear while loading or when the photo is unavailable. Outgoing replies show their text and status in the conversation.

For an outgoing Telegram reply, Accepted by Telegram means Telegram accepted the API request. It does not prove that the teammate received or read the message. If d5s cannot tell whether an interrupted request crossed Telegram's acceptance boundary, the activity remains Delivery unknown and d5s does not blindly repeat it. Check the private chat for the original before asking the agent to send it again.

Disconnecting Telegram revokes the d5s installation, teammate links, and direct-message bindings. It does not invalidate the bot token at Telegram. If the token must stop working, revoke or rotate it separately with BotFather.

Discord agent channel (beta)

Discord is a limited beta and appears only after d5s enables it for your workspace. Each agent uses its own Discord application and bot. An authorized Owner or Admin opens Manage agent → Channels → Discord, then:

1. In the Discord Developer Portal, create an application and add its bot.
2. Copy the Application ID, Public Key, and bot token into d5s, then select Connect Discord. Treat the bot token like a password; do not paste it into a conversation. d5s verifies that all three values belong to the same application and registers the global /ask command.
3. Copy the Interactions Endpoint URL from d5s into the application's General Information page in Discord. Discord verifies this endpoint with a signed request.
4. Install the application in the intended server with the bot and applications.commands scopes and permission to view and send messages in the channels you will connect.
5. Enable Developer Mode in Discord, copy the Server ID and Channel ID, and add that exact pair under Server channels in d5s. d5s verifies that the connected bot can see the channel before saving it.

In a connected server channel, use /ask and put the request in its message field. Any Discord member who can use the command in that channel can start work with the agent's configured capabilities, so connect channels deliberately and turn off unneeded project connectors. Ambient messages, mentions, attachments, bot messages, edits, reactions, and thread activity do not reach d5s. A server command enters the agent's continuing conversation and replies to the same channel as plain text.

For a DM, open the bot's Discord conversation and use /ask. The first command returns a private browser link. Sign in to d5s and confirm the Discord account shown; the link expires after ten minutes. The d5s account must still be an active workspace member with permission to view and run the agent. Then repeat /ask in the same DM. DMs use the same persistent agent session, and the agent can reply only to the current linked-DM turn; it cannot start a proactive DM.

Discord replies are split at Discord's 2,000-character limit and suppress all automatic mentions, including text such as @everyone. Accepted by Discord means Discord accepted the API request; it does not prove that a person received or read the message. Ambiguous partial or interrupted delivery remains unknown rather than being blindly repeated. The agent can also send plain text to an explicitly connected Discord channel through its normal destination tools.

When a Discord-started run pauses on an Ask-gated connector action, the agent posts Deny, Allow for this run, and Allow once where the work started. In a DM, only the same currently linked Discord account can decide. In a server, Discord must verify the click in the exact server channel bound to the agent. d5s rechecks the installation, binding, agent, workspace, and connector policy before accepting the first decision. An unanswered control expires after 15 minutes; an old, copied, or repeated interaction cannot grant access.

Disconnecting Discord revokes its saved channel and DM bindings in d5s. It does not delete the Discord application or invalidate its bot token. Remove the application from Discord or rotate the token there when access must stop outside d5s too.

Slack setup

The Shared d5s Slack app guide under Manage agent → Channels → Slack walks through installation, channel access, agent selection, and a first reply. Use Previous and Next to revisit any step. Connect Slack stays below the guide until the workspace is connected; Open Slack remains there afterward. The Dedicated identity section follows on the same page, and the Channels overview shows one Slack row whose status covers both paths. Connected guidance starts at channel setup, not a claim that your first reply has been verified.

Open a Slack channel’s details and select Agents & apps to add d5s. Mention @d5s with “Reply here with: Slack test passed.” If prompted with Connect d5s, follow the private link to sign in with your d5s workspace account and finish linking, then return to Slack. This personal account link is separate from installing the app. Select an accessible agent from the private Choose agent dropdown and check for the reply in the same thread. An existing thread selection can skip the picker.

The quickest Slack setup installs the shared @d5s app once for a workspace. A workspace admin opens Manage agent → Channels → Slack (the Set up action on the agent's Channels popover leads there too), chooses Connect Slack, and approves Slack's access screen. Other members see guidance to ask an admin. Once connected, choose Open Slack, invite @d5s to a channel, and mention it with your request.

If that thread has no selected agent and the channel has no default, d5s asks who should help inside the thread. Choose an available agent; the confirmation shows that agent's name and character, and the agent receives your original request so you do not have to repeat it. d5s remembers that choice for later mentions in the same thread. Another thread can use another agent. Change agent changes the thread choice; Set channel default is optional and makes that agent answer new threads that have no choice yet. If there is only one available agent, d5s can start with it directly. Shared-app messages still require an explicit @d5s mention. When an agent starts a new channel update, that thread is assigned to the posting agent: mention @d5s in its thread to continue with the same agent. You can still change the agent for later messages.

The picker shows agents available to you in that channel. Agents with the same name include context to help you distinguish them. Existing connected agents do not require you to repeat the channel setup. Slack OAuth links the installing user’s verified Slack identity to their d5s account. Other teammates whose accounts are not linked yet follow the private Connect d5s action first; d5s then resumes their waiting request. Once an agent is connected to a channel, channel members can ask it to use its configured shared capabilities. Replies from the shared app show the answer first, with the selected agent's name and small character in a compact footer and use Agent name · d5s as the native sender. A workspace admin must choose Reconnect Slack once for each connection installed before this identity permission was added. The visible in-message identity remains part of every reply so the agent stays clear when Slack groups messages. The optional @d5s +poppy shortcut selects an agent by its unique normalized name; it is not required and does not autocomplete in Slack.

For a genuine mentionable identity such as @Poppy, choose Create @Poppy under Manage agent → Channels → Slack → Dedicated identity. If Slack provisioning has not been connected yet, a workspace admin pastes a Slack configuration refresh token in that same setup panel. d5s connects it once for the workspace, creates a separate Slack app from the agent profile, and sends you to one Slack approval screen. Later dedicated coworker setup and permission upgrades reuse that workspace connection. If Slack revokes the credential, the action that needs it asks an admin to reconnect there and then retries automatically. Invite the dedicated coworker to a channel and mention it; the first mention connects that channel. Members of a connected channel can ask the agent to use its configured skills and the workspace connectors still enabled for its home project, so turn off unneeded connectors and connect channels deliberately. Disconnecting a channel under Manage agent → Channels → Slack → Dedicated identity stops the agent answering there; the app stays in the Slack channel until you remove it from Slack.

A Slack mention or linked direct message can include files. When you mention the agent in a thread, it can also discover files shared in earlier messages within its bounded thread context; you do not need to attach them again. If the thread or attachment limit is reached, the context reports that content was omitted. The agent downloads a file only when it needs it, including when it finds the reference later in conversation history. A downloaded copy remains available to that session, including after its working environment restarts; an unchosen file is not copied into the working environment. Treat downloaded files as untrusted channel input. The agent can also attach an artifact it created to a Slack reply. Each file and each message has a size or count limit, and a failed attachment is reported instead of being silently omitted.

For an outgoing Slack reply, Accepted by Slack means Slack accepted the API request. It does not prove that a teammate received or read the message. If d5s cannot distinguish a timeout after acceptance from a failure before acceptance, the activity remains Delivery unknown and d5s does not blindly post the message again. When Slack explicitly proves a temporary refusal before acceptance, d5s can make a bounded retry only after rechecking the current agent, channel, and originating access. Do not manually repeat an Unknown send unless you have first checked Slack for the original message.

Slack shows its native processing state while the coworker is working. Choosing Stop cancels only that exact Slack thread: queued work is removed before it runs, while matching in-progress work receives a graceful stop and replies with a short confirmation. Work in another bound channel or thread is not interrupted.

If Manage agent → Channels → Slack → Dedicated identity shows Upgrade permissions, a workspace admin must continue to Slack and approve the added permissions for abilities such as incoming files and thread follow. The existing app, connected channels, and conversation history stay in place. Until the upgrade is approved, the coworker uses its existing permissions and its first reply explains that thread follow is unavailable. That reply links to Upgrade permissions in d5s. Members who cannot manage the connection should ask a workspace admin.

Channel members also answer the agent's approval requests when Manage agent → Autonomy → Tool approvals is set to Require approval. When a run started in that channel pauses on a connector operation whose policy is Ask, the agent posts an approval card in the thread and any member of the channel can allow it once, allow it for that run, or deny it. An unanswered card is denied after 15 minutes. The same agent setting applies to every connected channel and to scheduled wakeups. Ask-gated actions in unattended runs remain blocked rather than creating a card that waits for someone to return. Allow without approval skips these Ask cards everywhere the agent works, while explicit Deny and capability boundaries continue to win. See Connectors for what an approval grants.

In a bound channel the agent can @-mention people on request. Mentions stay allowlisted: beyond whoever is already part of the conversation, the agent may list the channel's member names to work out who a nickname refers to, then deliberately look that person up by name — and only verified members of that exact channel can be resolved, with a small per-turn cap. People elsewhere in the Slack workspace are never disclosed or tagged.

After a mention opens a thread, the coworker follows that thread: human follow-ups in it reach the agent without another mention, so the conversation flows like one with a colleague. The agent's first reply carries a short footnote stating the contract; an app that still needs the permission upgrade gets the upgrade prompt instead, never a follow. It replies only when a message asks for its input and stays silent on acknowledgments and side conversation, though an explicit mention always gets a reply. A follow ends on its own after 72 hours of thread quiet, immediately when someone reacts to one of the agent's messages in the thread with :no_bell:, or when Thread follow is switched off for that channel under Manage agent → Channels → Slack → Dedicated identity; a fresh mention in the thread starts it again. Messages outside a followed thread are discarded before anything is stored — only followed-thread messages and mentions ever reach d5s.

Messages from third-party Slack bots never wake an agent.

Messages from d5s and connected channels use the same agent and continuing conversation. Slack, linked Telegram, and email activity remains attributed to its source and reply destination. The agent uses its configured capabilities and shared context across those surfaces, within their permissions.

Everyone who can open an agent in d5s can review activity from its currently connected Slack channels, even when their own Slack account is not linked or is not a member of those channels. This is read access to the shared coworker ledger, not Slack action authority: sending, replying, cancelling channel work, and other provider actions still enforce their own identity, channel, and live-membership checks. Connect private Slack channels only to agents whose d5s viewers should be able to inspect that channel activity.

Every agent-authored Slack message and native agent email visibly identifies the sender as an AI agent. Emails from a d5s-managed address use a compact signature with the agent’s avatar and AI-assistant identity. Where branding is enabled, Request access opens the verified recipient flow for colleagues, while Get started leads to normal signup for recipients who want d5s for their own team. Connected Gmail mail retains its existing AI identifier and application link. Paid email unbranding can remove the separate d5s footer and link, but it cannot remove the AI identity.

In the agent sidebar, select a row under Connectors to open the agent’s connector access settings. Select a row under Channels to open that channel’s settings directly.

Give an agent its own Teams identity

When Microsoft Teams provisioning is enabled for your organization, open the agent's Channels → Microsoft Teams panel. An organization owner or admin connects Teams provisioning once, then selects Create @Agent. The named identity uses the agent's d5s profile. Team members mention that agent directly; they do not need to mention a shared d5s bot or type an agent selector.

Creating an identity is separate from installing its app. Use Continue in Teams to select a team and finish Microsoft's consent flow. After installation, mention the agent in an intended channel and refresh its destinations in d5s. Activate the verified channel for that agent. One named identity can be used in multiple channels, and different named agents can share a channel.

Teams installs apps at team scope. Activating selected channels in d5s controls where that agent runs; it does not narrow the permissions granted by Microsoft. Follow replies requires the applicable Microsoft message permission. Use the permission-upgrade action and refresh permissions after approving in Teams. Only enabled channels and threads the agent has joined are admitted for follow-up turns. Otherwise, mention the agent again.

Sync Teams profile updates the existing identity from the agent's current profile. Removing a channel stops that route. Disconnecting an identity stops its Teams ingress and delivery while retaining d5s history; other agents are not substituted. If provider cleanup needs a retry, the panel reports that state. When the panel warns that a credential is nearing expiry, an admin can select Renew credentials without copying or entering a secret.

An agent's normal assistant response remains in d5s. Posting back to Teams uses its explicit messaging tool and the agent's existing approval policy. The current custom-app flow supports standard team channels. It does not claim private or shared channel membership support, Teams direct messages, or Teams file transfer.

Review calls and transcripts

When meeting attendance is available for an agent, select the phone in the agent's capability sidebar to open Calls. The panel separates History from Auto-join, keeps its header and tabs visible while the selected section scrolls, and pins History pagination below the list. While the agent is attending a call, the phone has a neutral active treatment and History shows the live call with its scheduled time. Ended calls remain in pages of 20 so a long-running agent's earlier meetings stay reachable. Select a call to read its transcript, title, time, and duration.

Calls also reports the meeting-entry lifecycle. Waiting for host means the listener reached the lobby and asked to join. If that request expires, History shows Not admitted; if the host refuses it, History shows Declined by host. Call failed means d5s has only a less-specific provider failure, which can occur before or after entry. These states explain what happened when a transcript is absent or incomplete instead of presenting the call as though nobody spoke.

Under Auto-join, review the connected calendars, joining rule, and upcoming meeting switches. An externally organized meeting can still require the host's approval, especially when the agent's meeting address is not on the invitation. Add the address shown in Auto-join as a guest when possible, and make sure the event also appears on one of the selected calendars.

After a transcript is ready, the agent adds a compact Call complete card to its continuing conversation with the call title, duration, and short outcome. View transcript opens that exact call in the Calls panel. Use Ask agent from the transcript when you want to discuss what happened or request the next task in the normal conversation.

The After each transcript setting in Calls controls what that agent should do with future transcripts. By default it posts a short outcome in the conversation, keeps the full transcript in History, and asks before taking action. Edit this instruction when you want a different analysis or handoff; processing the transcript itself does not grant new connector permissions or authorize an external action.

Agent email can keep a d5s address and a dedicated Gmail mailbox connected at the same time. Open Manage channels → Email and select the provider tab to edit that mailbox's incoming and outgoing address rules. New agents created by signed-in users with verified email receive an automatic d5s inbox where available, with organization business-domain or exact personal-address rules and approval required before sending. Manual mailbox setup starts with empty rules; a channel manager can edit the allowed people or domains, or explicitly permit anyone. Saved outgoing rules can authorize sends without individual approval. Blocked senders take precedence over incoming permissions, including the optional reply exception for people that mailbox successfully emailed. Existing policies remain until explicitly changed. Everyone who can open the agent can audit its incoming and outgoing email cards. Read Agent email before enabling it.

Runs remain separate

Direct messages, manual wakes, and scheduled wakeups all create turns in the same session. Each accepted turn still has an independent run lifecycle, tool activity, and cost.

Only one run works in the agent conversation at a time. Direct messages can queue or join the current work, and scheduled wakeups retry when the agent is available. This prevents overlapping work while preserving the schedule.

The daily budget applies to all spend attributed to the agent's home project for the current day in the agent's configured timezone, including manual and autonomous work and sub-agents. Reaching the cap blocks further work for that day; it does not permanently pause the agent.

Continue to wakeups and schedules for the exact trigger, retry, concurrency, and cost behavior.

Archive an agent

Archive an agent that you no longer use but want to keep. On the Agents page, open the ⋯ menu on the agent card and select Archive…. The same menu has Edit agent and Pause or Resume. Anyone who can edit the agent can archive it.

An archived agent stops work:

• Its scheduled wakeups do not run.
• Messages in its channels get no reply.
• Nobody can start new work in its conversation.

Its conversation, memory, files, and settings stay. You cannot archive an agent while a run is in progress. Stop the run first.

To see archived agents, select Show archived (the archive icon next to New agent). An archived agent card has Unarchive and Delete…. Unarchive makes the agent active again with its schedules and channels. An agent that you paused before you archived it stays paused. Delete removes the agent permanently.

When you open an archived agent, its conversation history shows with an Unarchive button. The message box stays disabled until you unarchive the agent.

Contributing to a dashboard

Select a workspace agent in a dashboard's Work with selector to ask it for a specific change. Each contribution starts with a fresh conversation and compact board context; the agent can inspect individual widgets and sources as needed. A completed contribution adds a dashboard link to the agent's home activity, without copying the dashboard conversation there. You can also ask an agent in its own chat to update a dashboard it is permitted to edit.

Save finished files locally

On supported macOS desktops, choose folders and grant an agent Read only, Write only, or Both access. The agent uses explicit file operations during your Desktop conversations; presenting an artifact does not automatically save it locally. See Work with files on your Mac.