An AI Agent is a defined worker: instructions, tools, and a reason to run.
Anatomy of an agent
- ›Instructions - its standing brief, in plain language: role, scope, what good output looks like.
- ›Tools - which capabilities it may use (read modules, draft content, propose actions). Fewer tools mean more predictable behavior.
- ›Schedule or trigger - run on a cadence ("weekday mornings") or on events ("when a deal enters negotiation").
- ›Autonomy - what it may do without a human: from suggest-only up to delegated execution within its contract.
Where agents live
Command Center, Agents lists yours; the marketplace offers templates to install. Runs cost about 5 credits each and appear in the Fleet.
Test runs
Every agent supports a dry run before you save changes: it executes against your real data but writes nothing, showing you exactly what it would have done. Use this after every instruction change - it is the difference between confidence and hope.
Operating agents well
1
One job per agent. "Sales assistant" drifts; "flag deals with no activity in 14 days and draft a nudge" performs.
2
Start at low autonomy and read its proposals for a week before delegating.
3
Review the Fleet weekly; retire agents whose output nobody uses.
Important: Agents act within the permissions and contracts you give them - an agent cannot read a module you have not granted it, and cannot exceed its rate or spend caps. See Authority contracts.