Skip to main content
Configuring an agent means deciding what it’s for, how it should behave, and which capabilities it’s allowed to reach for. You set a handful of core fields, then attach the tools and skills the agent needs. Everything you configure is saved with the agent and applied on every run.

​
Core fields

​
name
string
A label for the agent. It’s for you and your team to recognize the agent — it doesn’t affect behavior.
​
goal
string
What the agent is for. A clear goal keeps the agent focused across interactive, API, and scheduled runs.
​
instructions
string
How the agent should behave — tone, steps to follow, formats to prefer, and anything it should avoid.
​
model
string
The model the agent reasons with. Saved with the configuration so runs stay consistent.
​
knowledge base
string
The knowledge base the agent is grounded in. Chosen when you create the agent and fixed for its lifetime — see Deploying & versioning.

​
Toggles

​
web search
boolean
When on, the agent can search the web to supplement your knowledge base.
​
memory
boolean
When on, the agent can recall durable facts across runs. See Memory.

​
Capabilities you attach

Beyond the core fields, you choose exactly which capabilities the agent is allowed to use. Enabling them explicitly keeps behavior predictable and gives you control over what the agent can reach.
Integrations the agent may use — for example, sending mail or updating records in a connected service. An app must be connected and authenticated before you can enable it on an agent. AutoSage validates readiness when you save, reducing avoidable scheduled-run failures; reconnect an app if its authorization later expires.
Reusable authored procedures the agent can follow. Attach a skill in one of two modes:
  • Always — the skill is loaded on every run. Best for procedures the agent should follow every time.
  • Auto — the skill is loaded only when it’s relevant to the task at hand. Best for procedures that apply to some requests but not others.
Call your own HTTP endpoints as tools. Any secret headers you add are encrypted at rest and shown masked afterward, so credentials stay protected once saved.
Registered tool servers the agent can call. You select from the servers already registered in your organization.
Other agents this one may delegate to when a question falls outside its own knowledge. See Delegation.
Connect the apps you plan to enable before you save the agent. With the connection in place first, saving goes through cleanly and scheduled runs have everything they need from the very first run.

​
A good configuration

1

Write a specific goal

State the single job the agent does. A focused goal produces more consistent runs than a broad one.
2

Give concrete instructions

Describe the steps, tone, and output format you expect. Include what the agent should not do.
3

Attach only what it needs

Enable the connected apps, skills, and tools the task actually requires. Fewer, well-chosen capabilities keep behavior predictable.
4

Choose skill modes deliberately

Use always for procedures that apply every run, and auto for ones that only sometimes apply.

​
Next steps

Deploy your agent

Save your configuration as a new version.

Set up delegation

Let this agent ask other agents for answers outside its knowledge.