An individual AI bot with a primary objective and access to tools to help it accomplish that objective
Tools
Agents can call tools to make API calls to other apps or systems (i.e. lookup information in a database or send an e-mail)
Tool Call
A single invocation of a tool, including parameters that are passed to that tool (i.e. for sending an email, the agent would choose the send_email tool and also send a parameter containing the email content)
Model
The specific AI model that an agent uses (the model is like the agent’s brain), such as gpt-4 from OpenAI or Claude 3.5 Sonnet from Anthropic
Evals
Tests that are used to evaluate your agent, with an input prompt and an expected result
Embeddings
A mathematical representation of text as a series of numbers. Allows for computation such as “how similar is each chunk of text to my search term”
The more context you can share, the better. The agent can only be as helpful as the instructions you give it.Example:
Instead of saying “Find information on software version upgrade”, say “I’m helping a customer upgrade their version of [software name] because they finally have internal bandwidth to do the upgrade. Help me share the right instructions and links with them to ensure their upgrade goes smoothly.”
Focus on Positive Instructions
Don’t tell AI what NOT to do. Instead, be specific and targeted with your instructions of what TO do.
Share Your Objective
Agents can make multiple tool calls and adjust based on findings, but they need to know your overall goal, not just the first step.
Explore Capabilities
If you’re not sure what your agent can do, just ask! The agent can summarize its tools and suggest productive ways to use them.