Skip to main content
Building an agent comes down to a few choices: how it should behave, which model runs it, whether it gets its own command, and whether it can search the web. Make those choices, publish, and the agent is ready to use.

​
Give it instructions

An agent does its best work when its instructions are specific: what it’s for, what to include, what to leave out, and the tone to use. Write for the job — a call summarizer that always lists action items reads very differently from a deal-risk assessor. Keep the instructions focused on one job rather than trying to make one agent do everything. An agent also has a description; leave it empty and one is generated for you when you publish.

​
Choose a command (optional)

A command lets the agent be called directly by name. It’s optional — leave it off for an agent that only runs inside workflows. If you give it one, it must follow these rules:
  • No spaces.
  • 2 to 24 characters.
  • Unique across your team — no two agents can share a command.
  • Not a reserved name. ask-ai, configure, and help are reserved and can’t be used.

​
Choose a model

Every agent runs on one model, chosen from three providers:
  • OpenAI, the GPT model family
  • Anthropic, the Claude model family
  • Google, the Gemini model family
The model picker in the agent builder shows exactly what’s available to you. Which models appear there depends on your workspace: an admin can enable or disable individual models per provider, and the catalog changes as providers release and retire models. See AI Providers to manage providers, model availability, and your own API keys. Several models let you tune two settings: reasoning effort (how much the model works through a problem before answering) and verbosity (how much it writes, available only on OpenAI models). The range depends on the model: some start as low as minimal or none, and the top settings (xhigh and max) are there for heavy reasoning work. Pick the model first, then adjust what it offers. Match the model to the job, a quick classification and a nuanced deal analysis have different needs.

​
Turn on web search (optional)

An agent can use web search to look up real-time information — recent news about a company, current facts, anything beyond what’s already in your Context Graph. Turn it on when the agent’s job needs live information from the web; leave it off when everything it needs is already in your data.

​
Publish it

Creating an agent leaves it as a draft. Publishing is what makes it usable — a workflow can reference only a published agent, and a rep can only call a published agent’s command. Publish when it’s ready, keep refining in the draft afterward, and publish again to roll out changes. A draft keeps a change history you can review. You can also discard a draft to fall back to the published version, or unpublish an agent to take it out of service.

​
In practice

You build an “account researcher” agent: instructions to profile a company and flag anything noteworthy, a capable model with reasoning effort set to high, and web search turned on so it can pull in recent company news. You give it the command research, publish it, and it’s ready — both as a step inside workflows and as a command reps can call directly. Questions? We’re here to help at support@clearskies.cc