Agents
Create agents with their own instructions, models, skills and tools.
13 articles in this topic
- Create a custom agentNeda keeps retyping the same warranty and delivery answers, and Kian would like to give her a helper that already knows the shop's rules. Create an agent with its own name, instructions, model and tools.
- Edit an agentA few weeks in, Kian notices that the support assistant is wordier than a chat needs. Open the agent's editor to change its identity, instructions, model and every other setting, and check the result in the live preview before you save.
- Turn on agent behavior optionsNeda's customers want a quick answer, not an essay. Two small switches tell the agent to keep answers short and to name its web sources.
- Give an agent skillsKian has written down how Nilgoon's weekly summary should look, and he wants the agent to follow it only when someone asks for that summary. Give the agent skills, bundles of instructions it loads when a request matches them.
- Attach knowledge bases to an agentNeda's customers keep asking about warranty and delivery, and the answers are already written down in the shop's policy documents. Attach a knowledge base to the support assistant so it searches those documents in every chat and cites where each answer came from.
- Choose the plugin tools an agent can useKian has connected the inventory plugin, but the support assistant cannot use it until he says so. Give the agent access to your connected plugins, and limit it to the tools you want it to use.
- Let an agent delegate to other agentsOne agent that does everything gets muddled. Choose other custom agents in your current workspace that this agent can hand tasks to, so a stock question goes to the inventory analyst. Each one works with its own model, instructions, tools and settings.
- Configure agent memoryNeda should not have to explain a regular customer's order again every time she opens a new chat with the assistant. Choose how an agent keeps the current conversation, what it remembers about a user across conversations, and how it shortens long history.
- Protect sensitive data with guardrailsNow and then Neda pastes in a customer message that includes an email address or a card number. Guardrails detect personal information in what an agent reads and writes, and remove it, mask it, replace it with a code or stop the run.
- Make an agent answer in a fixed formatKian's order script needs an answer it can read the same way every time, not a paragraph of prose. Require the agent's final answer to follow a JSON Schema, so other programs can rely on its shape.
- Set agent execution limitsKian does not want one runaway request to eat the shop's plan. Control how an agent delivers answers and put hard limits on its runs, model calls and tool calls.
- Choose voice and vision models for an agentMany of Neda's customers send a voice message instead of typing. Pick the models an agent uses to turn speech into text and to create images. These are separate from the agent's main text model. The choices for speaking answers and reading text in documents are saved but do not yet change how the agent works.
- Connect an agent to a Telegram or Bale botAt eleven at night a customer still wants to know if the oven comes in black, and Neda has long gone home. Give the support assistant its own Telegram or Bale bot, so anyone can talk to it directly, day or night.