Last week, we hosted the Product Spotlight: Build agents in Tines, and it was a hit. We had so many questions that we couldn’t answer them all live, so we’re continuing the conversation here.
Before we jump into the Q+A, here’s a quick recap of the webinar in case you missed it:
In this session, Head of Product Stephen O'Brien introduced the AI Agent action and shared how it builds on our ongoing evolution of workflows. He talked through Tines’ commitment to flexibility – giving teams the power to run deterministic and agentic workflows in one secure platform. He also showed how to tailor agents to your needs using tools, templates, and chat modes.
For more on building agents in Tines and how we got here, check out this blog Stephen wrote at launch.
Questions
Here are a few questions we wanted to call out specifically:
How can you identify the proper use case for an AI agent?
When considering the job you want to use an AI agent for, consider the following breakdown of the difference between agentic and deterministic workflows:
What Tines tools are available to use with the AI Agent action?
The tools you add to your agent come with the full power of Tines. They can range from action templates, “send to stories”, or groups. If you want to use an agent with another separate agent, you can invoke a “send to story” that utilizes a different AI Agent action.
What are the best ways to kick off an agent?
The AI Agent action exists as part of the Storyboard and can be kicked off by any preceding action, page, or autonomously via a prompt.
Can you interact with an agent through Slack as an app or Slackbot?
Yes. You can use either chat or task mode to interact via Slack.
How does the chat mode of an agent differ from the Tines copilot offering, Workbench?
The agent’s chat is powered by the tools used to build it and can be accessed by end-users. It also includes a pre-determined ending point for the conversation. For example, end-users can interact with a chat version of an IT HelpDesk to solve a specific issue. The conversation will end at a determined point in remediation, and downstream automations can continue to run after.
Workbench, on the other hand, is a copilot meant for builders within Tines with specific permissions, and cannot be accessed by end-users. It’s intended to be an open dialogue between the builder and the LLM, with the builder determining when the conversation ends.
Can you control the amount of data an agent uses in certain contexts?
With Tines, you are in control of what data gets passed to the agent. It will only see and interact with the data you explicitly grant via the attached tools.
What AI models are available in Tines?
See the available models in our Docs here, including the ability to bring your own model. We update this page with the models as they become available..
Can you choose from various AI models to run within the AI Agent action, instead of using one provider for the entire tenant?
Yes, inside the AI Agent action, you control which model is used. We even offer the ability to quickly choose between faster or smarter models.
What is the input token limit for the AI Agent action?
With the built-in Anthropic models in Tines, the input token limit for the AI Agent action is 200K. The token limits can be higher if you choose to bring an OpenAI or Google provider.
Can you control the token limit without losing important threat context?
Yes, you can estimate the token count before using the AI Agent action. We suggest using the function ESTIMATED_TOKEN_COUNT to get an approximate number of tokens based on your input data. See examples in our Docs here.
How do agents handle errors or failed LLM outputs?
Agents are on the Storyboard, which means they are supported by Tines' current error-handling features, such as monitoring for error responses or retrying failed requests. Depending on the type of error, Tines can help mitigate or recover from it.
From our own experience, agents will sometimes take a different approach to an output than expected if they don’t have the full context. The agents you build should be given the latest information and up-to-date access to your proprietary systems to ensure they can correctly interpret your data.
Are there plans for Tines to produce AI agents specific to certain use cases (e.g., a “Phishing AI agent”, a “DLP AI agent”, etc.)?
With Tines, we’re taking a slightly different approach. Your organization's subtleties and complexities mean that you need flexibility. An off-the-shelf agent wouldn’t be effective for all organizations that need something laser-focused for their unique use cases. The ability to build your own agent creates more opportunities for you and your team.
Will sample stories by the Tines team be available for builders to use themselves?
Yes, all available stories can be found in the Tines Library, with workflows specific to building agents under the AI Agent action section.
Check out the full webinar to see Stephen’s demos and listen to the rest of the Q+A
If you’re interested in trying out agents for yourself, secure your spot at our upcoming agents bootcamp or request your free trial today.