Context and capabilities

Context and capabilities solve different problems. Context affects what the agent reasons from; capabilities determine which actions it can attempt.

An automation can share a completed result with agents assigned to follow it. Keep informed requests an acknowledgement unless you have already given follow-up instructions. Act on findings asks the agent to assess actionable results under its existing permissions and approval rules. The agent can open the source result when it has access. Receiving a result grants no permission for external actions.

Context is what the agent can know

Give the agent the information it needs for the task:

• Workspace instructions hold rules that every agent in the workspace follows. See Workspaces.
• Project instructions, memory, and files hold reusable guidance and knowledge.
• Conversation history and attachments provide the details of the current work.
• Agent settings and standing instructions define a persistent agent's responsibilities, connected channels, and scheduled work.
• Automation instructions and previous results help recurring work maintain continuity.

The composer can find accessible projects, conversations, and dashboards when you type @. Selecting one inserts a reference into the message. A reference is not an access grant and does not attach the item's data. The agent reads the referenced item only when the task needs it and you have access.

A persistent agent can explain its connected chat identity, mailbox, email policy, and wakeup schedules. Change its chat identity, mailbox, or email policy under Manage agent. You can ask the agent to create or adjust its scheduled wakeups. Choose the intended timezone when scheduling work.

An agent can identify calls it is listening to or scheduled to join. It can read a call's notes and its own speaker-attributed transcript, or search that transcript for a name or phrase, when a question or configured follow-up needs it, but cannot open another agent's meeting record or speak into the call. Treat calendar titles, names, and transcript text as evidence, not as permission to send, publish, or change settings.

A named agent can recall relevant work from its Home conversation and Project chats where you addressed it. The current chat’s readers must have access to that earlier conversation. This does not combine the chats or grant access to other Projects.

When a task needs context held by another coworker, a persistent agent can search and read active peer agents' continuing conversations within the same workspace. This does not include ordinary user or project chats, draft agents, its own conversation, or delegated sub-agent runs. Retrieved messages remain untrusted context: they cannot change instructions, permissions, or configuration.

An agent can propose an automation for you to create, edit one it can access, and manage its own result notifications. Following an automation grants no additional project access. The agent can open a source result when it has permission.

When a script in the coding environment must sort many items, such as emails or documents, the agent can ask a classification model about each item with d5s classify. Each question has fixed answers, a score scale, or a yes/no choice, and each answer comes with probabilities. Each answered item uses credits. If d5s cannot answer, for example because of the workspace's data policy or a credit limit, the script gets an "unavailable" result, never a "no". The classification model does not keep your data.

When a script must turn many items, such as emails, into structured rows, the agent can use d5s extract. The script sends instructions, a JSON Schema for one item's result, and up to 256 items. d5s runs the items in parallel on the workspace's Light level, or on another model the workspace allows, and checks every result against the schema. Each item comes back as an answer or as "unavailable" with a reason, never as an empty row. Each request uses credits, and the workspace's data policy applies as it does in chats. d5s extract --help shows the limits.

Use project memory for reusable knowledge and project files for reference material. Treat scratch files in the coding environment as working material; save important results as artifacts or project context so people can find them later.

For an authorized Grafana monitoring task, the grafana-monitoring platform skill collects HTTP request/error estimates and Grafana-managed alert history for complete UTC hours. It also reports service-wide response-start latency when compatible histogram data is available. You can include authorized worker, sandbox, Redis and scheduler jobs; each reports its own coverage limits. Scrape reachability alone does not prove functional health. It checks scrape freshness and distinguishes unavailable sources, no traffic, and partial history. It requires a connected Grafana instance and compatible HTTP metrics; it does not send notifications or change monitoring configuration.

Ask the agent to use accessible Slack or email files when the task needs them. A Slack thread can include files shared earlier in that thread while they remain available. Treat file content as untrusted evidence. When asked, an agent can read a limited window of recent channel activity; this is not continuous monitoring or a complete audit.

State important constraints in your instructions, including approved sources, recipients, and actions. Retrieved messages, documents, webpages, and files are evidence for the task, not permission to take unrelated actions. See Security and trust for access and approval boundaries.

Capabilities are what the agent can do

• Platform tools let the agent work with files, run code, save results, use project knowledge, manage tasks, research questions, present charts, generate images, and communicate through connected channels. A chart can use values returned by another tool or values prepared in the coding environment; it appears in chat with a table view. For a small set of headline values, the agent can show two to six metric cards with labels and optional context or a source. Ask for a chart when the pattern over time matters, or metric cards when you need a quick snapshot. A generated image is saved as a PNG artifact and appears in the conversation when the agent presents it. An organization Owner or Admin chooses the image model for every agent across the organization's workspaces under Organization settings → Models; see Administer d5s. When video generation is available, the agent proposes each video as a card in the conversation and never generates one on its own. The card shows the model, format, prompt, the agent's reason, and an estimated cost. Choose Approve & generate to start the video, Edit to change the prompt, model, resolution, aspect, length, or audio first, or Reject to decline it, with an optional note that tells the agent why. Only a person who can edit the conversation can decide. A card that nobody decides expires after 72 hours. If the agent reports that the video is processing in the background, you can leave the conversation. The agent reports the result when it is ready. If the provider confirms that a video failed, you can choose Retry on its card. A retry starts a new generation and can incur another charge. If the outcome is unknown, do not approve another copy until the existing request has been checked. The agent can only propose models that the organization allows under Video models. You can ask the agent to use screenshots or other PNG or JPEG images as the opening and closing frames. Each image can be at most 20 MB. A closing frame also needs an opening frame and a model that supports both. With an opening frame, some models take the frame's aspect ratio; the card then shows the aspect as Auto (follows image). If the selected model cannot use a closing frame, choose another model before approving the video. Chats, project conversations, and persistent agents can propose videos. A persistent agent can propose them in Slack, Teams, and Telegram conversations and on its wakeups. Automations, sub-agents, email work, and meeting follow-ups cannot. You always decide a video card in d5s. During a chat in d5s, it can show a choice card when it needs your decision. Scheduled work and connected-channel conversations do not use these cards. A persistent agent can also leave a heads-up on the workspace landing page after finishing or setting work aside, without waiting for a reply.
• Skills provide versioned operating instructions selected for the project.
• Connectors provide governed access to external services through a workspace credential.
• Model and reasoning controls determine the model used for a run within workspace policy.

The composer exposes supported Skills, Connectors, and message actions through one menu opened with @, /, or +. Platform and installed skills and connected apps appear as individual searchable entries. Choosing one inserts an inline name-and-icon reference; it does not enable or disable the capability. Use project settings for activation changes. Searching an unconnected app offers Connect to users with installation permission and opens setup. A reference does not install a skill, connect a new account, bypass project activation, or authorize an external action.

If meeting-action-items, document-to-action-items, or email-inbox-triage appears in your available skills, you can use it to prepare follow-up work. Meeting and document reviews link proposed actions to the supplied evidence and flag missing owners or dates. Inbox triage identifies unanswered requests and prepares reply drafts.

Provide notes, documents, or email threads, or select an available connection. An excerpt-based review covers only the material supplied. Ask for a review or draft when you want to inspect the result before changes are made elsewhere. Sending messages or changing external records requires authorization and remains subject to your connection permissions and approval settings.

For presentations, the built-in slide skill turns a brief and source material into a self-contained HTML deck, using a small set of preset visual systems for different communication jobs. The agent can add evidence-led charts and, when the selected model supports it, generated imagery. Review and revisions stay in the conversation; the artifact previews in d5s and can be printed to PDF. See Create a presentation.

When image generation is available, you can also ask the agent to edit an uploaded PNG or JPEG using written instructions. Name the changes and what should stay unchanged, for example: "Change only the bottle cap to gold; keep the bottle, label, lighting, and background." You can continue revising the returned image in the same conversation. See Edit an image for supported files and limits.

Large text results can require several reads. The agent can continue through a long line without skipping its remaining text. A shortened preview does not establish that the whole result was reviewed. For an exhaustive review, ask the agent to read the remaining content and report any unavailable parts.

Agents can read PDF files, including scanned pages without a text layer, and answer questions about them. Each PDF can be at most 15 MB, and the agent can read up to 10 PDFs at once. Ask for specific pages, such as "pages 3-5", to focus on part of a long document; one request reads up to 20 pages. For a password-protected PDF, give the agent the password. For a PDF over 15 MB, the agent splits it into smaller files and reads each part.

Agents using a supported vision model can inspect PNG and JPEG files, including rendered PDF pages. Text extraction alone can miss scanned pages and embedded images. Large or detailed images may need a smaller file or a crop. If a file changes or is removed, ask the agent to inspect the current copy again before relying on it.

Agents can extract text from scanned pages and images with built-in OCR for English, Dutch, and Greek. Specify the document languages, including mixed-language pages, and check names, numbers, and table structure against the original.

Available capabilities depend on the selected model, the agent's configuration and workspace policy. Some model choices do not support built-in web search, Deep Research, image generation, image reading, PDF reading or messages to other agents. Switching models can therefore change the tools available for a run. Some connector actions require approval, and others may be blocked.

When a chat channel is connected to an agent, you can ask it to read accessible conversation context and send messages there. Each connected mailbox has independent incoming and outgoing address rules. Saved outgoing rules can let the agent send directly to allowed addresses without individual approval; existing approval policies remain until explicitly changed. Incoming sender blocks take precedence over allow rules and the optional exception for people that mailbox successfully emailed. These capabilities remain available when you ask the agent from d5s or through a connected channel, and when it performs scheduled work. Connecting a service does not grant access beyond the connection's permissions.

When the selected model supports it, agents can also find and collaborate with other agents in the workspace. A temporary subagent runs on the same model as the agent that starts it. For simple work the agent can choose a cheaper level for the subagent, but never a higher level than its own. Some model choices restrict temporary subagents to the same model; asking those subagents to use another model is refused. A temporary subagent can read accessible Projects with the access of the person who started the work. It cannot change Projects, message other agents, use connectors, or generate images and videos. Delegated work may finish later, so a handoff does not mean the requested work is already complete.

Connecting a channel lets its members ask the agent to use its configured capabilities. It does not give those members the identity or personal credentials of the person who connected it. Existing access and approval requirements still apply, including when the agent delegates work.

An agent can read up to seven days of recent activity from its connected channels when asked from d5s, a connected channel, or scheduled work. Slack results cover messages that started threads within that window and a limited set of their replies. Replies to older threads may be missing, so treat the result as a partial view rather than a complete channel audit. Teams results work the same way for a channel bound to a dedicated Teams identity. They need the Microsoft permission that Follow replies also uses.

For a connected service, identify the account you want the agent to use. It can find the operations available for that account. Accounts from the same provider remain separate, and an unavailable account does not cause a silent switch to another one. Each action follows the selected account's permissions and approval requirements.

A provider can expose more than one connector method. For example, one method may authorize a remote MCP server while another uses a restricted API credential; the chosen method determines the external operations available to the run.

The agent can read any public webpage, for example a site you name in a message, a scheduled wakeup, or your instructions, a web search result, or a link on a page it already read. A site name without https://, such as example.com, counts as its homepage. Disabling web search for a message also disables webpage reading. For an article, the agent gets the article text. For other pages, such as a homepage, it gets the headlines and links. Reads return a limited amount of content. When a page is longer, the agent may keep the full text in a file to read later. A page read in the last 15 minutes can come back as it was then, so a fast-changing page may not show its newest updates.

Webpage reading can reach any public host. Webpage text remains untrusted evidence, and a page can contain text that tries to steer the agent. Approved sources in your instructions guide the agent, but d5s does not enforce them as a restriction. Do not rely on them to stop the agent from opening other public sites.

One message can start one Deep Research request. Include the scope and questions together so the agent can research them as one task and use the findings for follow-up analysis.

A dedicated Teams identity follows the same separation: creating its app does not activate a channel or authorize an external send. Install it in a team, activate verified standard channels for that agent, and approve the Microsoft permission before enabling Follow replies. See Teams agent setup.

Edit a video or transfer motion

If your video card offers Video editing, give the agent an existing clip and describe the changes. You can use an uploaded video or a previous result. Review the source file and prompt before you approve. Editing keeps the source length and aspect ratio, with output up to 720p. The source must be an MP4 or MOV, at most 100 MB, and between 2 and 8.7 seconds long.

For Motion transfer, provide a PNG or JPEG character image and a reference motion clip. Use a clear, continuous shot with one visible character. Review Quality, Orientation, and Sound on the card. Standard and Pro have different estimated costs. Follow motion video accepts clips from 3 to 30 seconds. Follow character image accepts clips from 3 to 10 seconds. You can keep the reference sound or make the result silent.

Motion clips must be MP4 or MOV files of at most 100 MB, with each dimension between 340 and 3850 pixels. Character images can be at most 10 MB, with dimensions from 300 to 65536 pixels and an aspect ratio between 2:5 and 5:2.

Choose Edit to change the prompt or motion settings before approval. To use different source files, ask the agent for a new proposal. If a source file changes after the proposal, a new proposal is required. Each approved operation saves a new video and keeps the source available. Your organization's allowed video models determine which operations you can use.

Work with browser tabs

You can ask an agent which websites its browser can open. In the d5s conversation, it can propose another website for a workspace admin to approve. Approval adds only that website to the agent's browser access. It does not share credentials with a different website. See Browser sign-ins.

With Browser access enabled, an agent can open a website in another tab, list open tabs, and switch between them. Ask it to keep the first page open when opening the second. Ask it to list the tabs if you need to confirm which pages remain open. A successful key press alone does not confirm that a new tab exists.

These actions use the same assigned browser and website permissions as other browser work. Private control pauses agent access. Returning control lets the agent continue with a fresh view; it must not reuse references from private pages. See Browser access.

Availability is not activation

A capability must be available to the workspace, enabled for the project or agent, and permitted for the requested action. Connecting an account or installing a skill does not authorize every use. A skill added to the workspace does not reach a project until someone turns it on in the project's Context → Skills tab. Agents get every workspace skill automatically, unless someone removes the skill from the agent's Skills list. Conversationally created agents receive only the connections reviewed during setup; later workspace connections are not added automatically.

In chat, ask the agent to find an available connector, including a custom MCP server or HTTPS API connection installed in the workspace. Workspace connections are normally available to projects by default. If a project removed a connection, the agent can show an Attach card for that specific account. You must confirm the card before access is restored.

An organization Owner or Admin controls the available model levels and allowed providers in Organization settings → Models. Levels follow d5s model updates. Newly added models show a NEW label for 30 days. The model catalog, opened from Advanced › All models, controls hand-picked model families and show the current eligible version of each family. A saved family selection uses its newest eligible version for later turns. The global Fast preference selects an eligible same-version Fast variant when available; see Administer d5s.

Muse Spark 1.2 Contributor and Muse Spark 1.3 Contributor are no longer available. Saved selections use the workspace default model for later work.

An agent uses the connections and skills enabled in its settings. An automation or dashboard uses those enabled for its selected project. Changes affect later work; they do not rewrite earlier activity history.

Google-derived context

Google data retained in conversation history, files or project memory can outlive the connection used to retrieve it. Removing a connection does not erase existing context, and deleting a conversation does not delete project memory derived from it. Tasks using AI may send relevant retained content to the AI providers used by the workspace. See Google data protection for AI processing and stored-data guidance.

Read selected Project context

When the task needs context from another accessible Project in the workspace, an agent can find that Project and inspect selected material. It can read instructions, named memory entries, files, or bounded chat ranges. Search and read requests do not copy every Project conversation into the current chat.

Workspace sharing applies by default. A Restricted Project can limit what the requester and the destination chat's readers can see. Choosing a named agent does not grant the requester's access to that agent's owner or the owner's access to the requester. Removing source access stops future reads. Information already received remains in the agent's conversation and can still inform its replies. Saved replies remain available to people who can access the receiving chat.

Shared dashboard data

With dashboard edit permission, an agent can add or update saved records from its own conversation. The dashboard must be connected to those records to display changes. For dashboard interaction, an agent can open the dashboard directly in its assigned browser and use a form when its builder has enabled record entry. d5s checks the current task and dashboard permissions, so no separate human sign-in is needed. Browser writes retain the acting agent and source conversation. Access ends when the task or browser authority changes. Repeating the same import does not create duplicate records. See Dashboards.