In version 8 of our platform we have introduced a powerful new AI feature called Agents. Agents are chatbots that can "actually do things". They can look up information, answer questions, provide details explanations and even perform actions on your behalf.

There are several built-in agents that specialize in different areas of the application, but you can also add your own agents, specific to your particular area of business.

In this section we'll detail the different features available to AI agents, to help you make the most of this new exciting feature.

Please note, though, that this is a new feature and as we're still ironing out the wrinkles, you should use extra caution when using the agents for anything important.

With the initial rollout, we are limiting access to agents to admin users. In the future we will expand this for Wiki Managers and other users, based on feedback while treading very carefully to ensure maximum security and control.

The Agents Tab

You'll find your agents in the "AI Agents" tab, visible for admin users. You'll also be able to access agents when using our platform, as a chat box in the bottom right hand corner. For example, when on an incident, you'll see an "Incident Reports Expert" box in the lower right corner where you can discuss the incident report you are viewing. The title of the chatbot will reflect which agent you are talking to, and what expertise it has. Any conversations you have in the chat box in the bottom right corner will also be available in the Agents tab under the relevant agent.

Agent Skills

Agents have a list of skills that you can manage. These skills are simple how-to text entries in a "Skills" list that tell it how to perform different actions or to go about answering specific questions. Some of the agents have pre-loaded skills, but you can also add to those skills yourself. You'll find this in the "Skills" tab under an agent, which you can access as an admin on the Agents tab. Note: for self-improving agents, this list may include skills the agent added to itself as it learnt. These should be reviewed regularly. You can revise these or remove them as you see fit.

Web Browser

Each agent has access to its own hidden web browser and has been trained specifically on interacting with the platform using the browser, but can also be trained on using other web sites. 

The user can also view the agent's browser at any time by clicking the "Access Browser" button. This gives you a "remote view" of their browser, which also allows limited interaction and keyboard entry (albeit, slowly). This remote view can be used, for example, to login to sites so the agent can perform work on them. Note that the height of the agent's browser is significantly taller than a regular browser, so it can view more information at once - this means you may need to scroll down to see any UI boxes that are centered on the screen.

Note: For on-premise users, the helper tool "cfagentbrowser.exe" must be running on the system, or installed as a Windows service. Chrome or Edge must also be installed on the server, with adequate memory available to support multiple tabs being used. For more information about this tool, please read this.

For maximum security, agents can easily be configured to have no access to a browser, access only to your wiki site, access to white-listed sites, or access to any site. You'll find this configuration option in the Settings page for each agent (accessible through the cog icon on the agent page).

However, note that the agent's browser will retain website cookies in the context of that specific conversation, and so access to the conversation by an admin user may allow access to the same site cookies if they have not yet expired.

Agent Actions

Agents can actually take actions on your behalf, whether updating a policy or collecting information for an incident investigation. They will have access to all of the actions/buttons that you can access on the site. When asked to perform a complex task, they will create a plan first (which is shared with the user) and then work through the plan until it is complete.

Creating and Updating Policies

As part of our platform, agents have special abilities to read large policies and make updates to them. If you ask an agent to update a policy, once in the editor it'll spawn sub-agents to summarize the policy contents and find the exact place it needs to update, remove or insert. This allows agents to process large policies without being too concerned about limitations in context-window size. 

Self-Improving

Agents are self-improving. If they are asked (or asked to do) something they are unsure of, or unable to complete, they will open a side-chat with their supervisor to ask for guidance - notified via email. This will be hidden from the user and only accessible to the supervisor. Once the answer is given, the agent will not only continue the task/chat, but will automatically update its skills so it knows how to handle that question or task the next time.

In-Training Verification

Agents that are marked as "in training" will always ask their supervisor to check (or correct) any forms/updates they fill out before submitting. For example, if a task it has to close an incident, it will bring up the form but will not click OK until it has checked the action with its supervisor.

Agent Collaboration

If an agent is asked to do something outside of its expertise, it will confer with another agent that it knows has more understanding of that area. it does this by starting a side-chat with the agent (which cannot be seen by the user). Once the agent has the information that it needs, it will then go back and continue the original discussion armed with that new information. 

This agent collaboration allows for optimal use of the available context-window size which can save costs. Because the side-chat has a smaller chat history, its context-window size will be smaller and more focused. This eases the burden on the original chat, which may be longer, because it can determine part of the answer without having to feed in the entire chat history of the original chat for each turn of the conversation.

Images and Infographics

Agents can create images and infographics to help provide explanations or guidance. Asking an agent to create an image will show the image in the chat window and store it as an attachment under the chat record. Likewise, you can ask it to create an infographic that will explain a certain concept, which can be copied and pasted into other documentation.

Note: Image and infographic creation is not available in the "Questions" box chat agent. 

Document and Spreadsheet Generation

Agents can produce documents and spreadsheets that can be downloaded. For example you can have it analyze some data and produce a spreadsheet, which you could download and vie win Excel. Or you can have it generate a Word or HTML document and download that also. These will be stored in the attachments area under the agent conversation.

Understanding Documents

You can upload documents and spreadsheets and reference them in your conversations. You can do this using the "+" icon in the chat window (this feature is not available in the Questions box chat bot). Once uploaded, the agent will analyze the document to help you answer questions, or even generate a new, modified document to replace it, as a download.

Agent Assignments

Agents can be associated with a user login that is specific to them. You set the associated user in the "Associated User" option in the Settings window of the agent. You can have it create a user itself by going to What's Next -> Create User For Agent.

Any assignments made to that user will be processed automatically by the associated agent in the background. 

You can control this assignment processing by editing the Background Processing section of the agent's settings window, which will dictate how often it checks for assignments and how long it will spend processing the assignment. You can do this through three settings in the "Background Processing" section of the agent settings: Heart Beat Minutes: will spend some time processing outstanding assignments every N minutes after it last checked. Max Tasks: How many assignments it will try to work on at once (any other assignments will wait until an assignment is complete). Max Turns: how many conversation turns it will allow before "going to sleep", waiting until the next heart beat. This allows you to control how much resource it will spend on processing assignments, as a cost control measure.

Security

Every conversation with an agent will only be accessible by the user and the agent, except that all admin users can see all conversations.

Because of security concerns, agents that can perform actions or access the browser can only be accessed by an admin user. Agents can be set to record a screenshot after each of the clicks performed by an agent, for auditing purposes.

Cost-Saving Features

Cost saving features let you control the level of reasoning and context-window usage allowed on each agent, along with setting maximum conversations per user per day, and maximum number of turns in a conversation. You can do this through two means in the "Advanced" section of the agent settings: 

• Context Portion: Specifies the percentage/portion of the maximum context-window size to use for this agent. For example, if the maximum context-window size for a given model is 1 million tokens, then setting this to Half will make it only use 500,000 tokens at a time. Because long conversations use more context-window size, and the LLM pricing is based on per-token costs, setting this to a lower value can help control costs.
• Max Turns: This is another way that you can control how long a conversation gets. By enforcing a maximum number of conversation turns (a turn is considered a single chat entry in a conversation), you can stop a conversation from getting too long - forcing the user to start a new conversation instead.
• Max Conv Per Day: This lets you control how many conversations a single user is allowed to have with that agent on a daily basis. Once they have created that number of conversations, any attempt to start a new conversation will be met with an error message. They can still go to other agents if the limits allow.