Agent Core
On this page
Purpose
Agent Core combines one input Message, instructions, its own earlier turns in the active session, a selected model, and optional tools. The model decides whether to call the connected tools before it responds.
When to use it
- Build an assistant that can calculate, retrieve data, or call approved tools.
- Run a bounded multi-step task where the model chooses the next tool call.
- Use Language Model instead for one response with no tool selection.
Required setup
Arkus is the default provider and needs no provider setup. Anthropic, Google Generative AI, and OpenAI require their matching API key.
How to use it
- Connect Chat Input to Input.
- Write instructions that state the task, limits, and expected output.
- Select a model provider and model.
- Connect only the tool-mode components the agent may use to Tools.
- Connect Response to Chat Output.
Configuration
| Setting | What it controls | Recommended starting point |
|---|---|---|
| Input | The Message the agent responds to. | Connect Chat Input. |
| Tools | The tools the agent may call. Accepts several connections. | Connect only the tool-mode components the task needs. |
| Agent Instructions | The instructions given to the agent. | Replace the default with the task, limits, and allowed tools. |
| Model Provider | Selects the model and its provider-specific fields. | Start with Arkus. |
| Model Name | The model the agent uses, listed for the selected provider. | Keep the default for the provider, or pick another listed model. |
| Temperature | Controls response variation. It is an advanced setting, and it is hidden for models that do not accept it. | Keep the default: 0.7 for Arkus, 0.1 for OpenAI, Anthropic, and Google Generative AI. |
| Max Tokens | Limits the tokens the model generates per response. Google Generative AI labels it Max Output Tokens. It is an advanced setting. | Leave it unset (Anthropic starts at 4096) unless responses must stay short. |
| Number of Previous Turns | How many of this Agent Core component's own earlier turns in the active session it can see. Each turn holds the input this component received and the response it produced. Another Agent Core component in the same flow keeps its own separate turns. | Start at 100. |
| Current Date | Adds one Current Date tool for each run. | Leave on when date or time affects the answer. |
| Handle Parse Errors | Lets the agent recover from malformed tool output. | Leave on. |
| Verbose | Enables detailed execution logs. | Leave on while building. |
| Max Iterations | Caps the agent's tool and response cycles. | Start at 15. |
Expected result
Agent Core returns one Message. Its observable Agent Steps contain input, tool activity, timing, errors, and output records. They do not expose private model reasoning. This component's own earlier turns in the active session load automatically, and Current Date adds a fresh tool on every run when enabled.
Reference details
OpenAI JSON Mode is not available in Agent Core because it conflicts with tool calling. Agent Description [Deprecated] is an advanced setting that applies only when this Agent Core is itself connected as a tool to another agent. It is deprecated, so leave the default. Connect only the tools needed for the task, then lower Max Iterations for tightly bounded work. Deployment availability depends on the components connected to the agent. Check the Agent Deployment guide before deploying.
Provider credentials and credits
Arkus uses Arkus credit metering and needs no key from you. OpenAI, Anthropic, and Google Generative AI bill the provider account behind the key you supply. For a deployed flow, that key must be a saved Credential rather than a value typed into the node.
Related components
- Chat Input supplies the conversation Message.
- Chat Output displays the response.
- Calculator and Current Date are useful bounded tools.