We make AI agents

How do I create an AI Agent?

AI agents are created and hosted on our platform. This platform is essentially an operating system for AI agents, and is continuously evolving and improving based on real-life experience working with customers to create useful agents.

Our platform is designed to make agent setup simple and intuitive, while enabling advanced behaviour by combining simple tools and instructions.

An example - Freddie: The Autonomous Invoice Router

Meet Freddie — an AI agent built to handle the busywork of invoice management with speed and precision.

After being briefed on company policies and key contacts, Freddie processes incoming invoices, flags any irregularities, and routes them to the right departments.

Every action is logged, ensuring full traceability and compliance.

This demo showcases how autonomous agents like Freddie can streamline operational workflows while keeping humans in the loop where it matters most

1. Create an agent.
2. Discuss your problem/need with the agent. It will suggest how it can help you. It will also suggest which capabilities/tools it needs to do the job, and will explain how to enable those.
3. After your onboarding conversation, the agent will write its own instructions, which you can review and tweak as needed.
4. Go back and chat or tweak the instructions at any point to fine-tune the agent's behavior.

This is similar to how you would work with a newly hired human colleague.

Note that the self-service approach typically works best for relatively simple use cases with well-defined processes. For more advanced workflows, we are happy to assist in co-creating the agent and making customizations as necessary (see below).

Co-creation workshops

For more advanced or customized workflows we offer workshops to help design and configure your agent. This is what we did when creating a journalist agent for the Swedish TV documentary "Generation AI" (see this demo video).

Co-creation workshops are useful when any of these factors apply:

• The agent's job is non-trivial and will involve multiple different tools.
• The agent needs to work with sensitive data or integrate with your systems.
• The agent needs to work with large amounts of data, and needs to be optimized.
• The agent needs new capabilities or features that we need to add to our platform.
• The agent needs a custom user interface or other specific adapations that require coding on our part.

The format of the workshops will vary depending on your context, but here is a typical setup:

• Workshop 1: Process mapping. We map out all the steps in the selected workflow, decide which part the agent will do, and which tools, integrations, and data it will need to do the job.
  • Between workshops: You gather the necessary information, data, and examples that the agent will need. We do technical work such as testing new API connections.
• Workshop 2: Agent prompting. We co-author the agent's instructions and iteratively improve it until we see that the agent successfully carries out the task.
  • Between workshops: You continue testing and refining the agent instructions. We implement any tools & capabiities & custom user interfaces as needed.
• Workshop 3: Pilot launch. We launch the agent in production, but perhaps with a limited scope. We test it together and monitor the quality of it's work.
  • Between workshops: You keep using the agent, validate that it is working as expected, and make minor adjustements in the instructions as necessary. We optimize the code as needed, and measure results.
• Workshop 4: Final delivery and expansion. We onboard more users, and make plans for further improvements or new agents.

Here is an example of a simple workflow plan that might come out of the first workshop. For each step we would need to figure out what information and tools the agent needs, what the input and output is, etc.

‍

How do I interact with my agent?

Our agents are designed to work like a human colleague, and like with any colleague you need a way to give instructions and communicate with each other.

To manage your agent, you log in to our platform and:

• Give it instructions. This is essentially a job description. For example "Every morning a 7am, check the latest news about climate change and send a summary to slack channel #climate-news", or "Whenever you receive in invoice, match it to an order number and email it to the appropriate team".

• Give it access to capabilities and tools such as slack, perplexity (a research service), email, hubspot, etc. It can also interact with your internal system, if you want it to. We are constantly adding new capabilities.
• Chat with it on the agent web portal, for example to ask questions or discuss and update its instructions. You can talk to it naturally using voice rather than typing. If you connect the agent to your slack, trello, or other communication tool then you can chat with it through that tool as well.
• Give it access to documents, for example an FAQ document, a price list, a contact list, or whatever information it needs to do its job.
• Read its diary to see what it has been up to.

‍

How do I ensure safety?

An AI agent is very powerful tool, and we have a huge respect for the importance of safety and control.

When designing an agent, we think carefully about:

• What is the scope of job? Does this agent have a very specific, isolated job (such as "add labels to a bug ticket"), or a more broad job (like "respond to customer inquiries")?
• What tools and data does the agent need to have access to? If it only can add labels to a bug ticket, then the worst case scenario is that it adds a bad label. But if the agent can do many things like sending emails, updating databases, accessing sensitive document - then we need to be more careful.

When designing agents we apply the principle of least privilege, which means we only give the agents the tools they need to do the job.

This is similar to hiring a human intern. If the job is to route invoices to the right department, the intern does not need to be able to send email or access sensitive data. As the intern "proves itself", you can give it more responsibilities and tools.

This is fundamentally a risk/reward tradeoff. An agent with a broad responsibility and broad set of tools and capabilities can be incredibly versatile and useful, and can handle complex tasks. But it also needs:

• Better AI models
• More carefully designed prompts/instructions
• More testing
• More guardrails
• More monitoring

Our platform has some built-in guardrails already, and we are constantly expanding this.

• We mostly use premium AI models from companies like Anthropic and OpenAI, which have a lot of built in safety measures
• Our platform includes watchdog agents internally that monitor the behaviour of all agents to make sure that they are acting in a way that aligns with their instructions and don't cause accidental damage.
• The agent's instructions are visible and editable by you, so there are no surprises.
• The agent writes diary entries after each action, giving you insight into what it is doing and why.

For important tasks that include sensitive data we recommend co-creation (where we help you make the agent) rather than self-service. That way we can work together to make the agent as useful as possible while ensuring safety.

‍

Agent teams

Our platform supports the creation of Agent teams. Similar to human teams, this is a collection of agents that communicate with each other and collaborate on different parts of a workflow. Like humans, agents work best when they have a pretty clear and well-defined responsibility. If you have a complex workflow, it is usually better to have several agents handle different parts of it in a well-defined way, rather than having a single super-agent with a very wide set of instructions.

Here is an example of an Invoice Coordinator agent who manages an invoice workflow, but delegates specific task to 3 more specialized agents: an invoice collector, an invoice analyzer, and an invoice router. Each specialized agent has its own specific instructions and tools, making it easier to configure and test their behaviours. In our platform you can connect agents in simple way like this.

‍