Custom Agents

Who Uses It

• Teams with repeatable research or reporting workflows
• Operations teams with multi-step back-office work
• Approved users in local or managed agent environments

Inputs

• Approved workflow instructions
• Secure CDM records
• Private Skills
• Source systems and public sources when approved
• Human review feedback

Processing

• Plans and executes multi-step tasks
• Reads approved context
• Uses approved tools and skills
• Pauses for human review where needed
• Creates structured outputs for downstream use

Human Controls

• Human-in-the-loop review
• Approved skill and tool access
• Permissioned data access
• Operator-visible outputs and source notes

Security and Privacy Notes

• Can run locally or in shared sandboxed environments based on customer needs
• Should use approved access paths rather than broad data exposure
• Sensitive actions should preserve review and audit trails

Outputs

• Research briefs
• Reports
• Structured workflow artifacts
• Recommended actions
• Approved updates through workflow APIs

Integrations

• Private Skills
• Agent Manager
• Secure CDM
• Research Agency
• Customer-approved tools

Implementation Pattern

• Start with a workflow where the inputs, output format, and review standard can be made explicit.
• Build the agent around approved tools and source rules.
• Add review gates before sensitive conclusions or actions.
• Measure whether the workflow becomes faster, clearer, or more reliable.

Results or Expected Outcomes

• Repeatable AI-assisted work instead of one-off prompting
• Clearer source trails and output standards
• More practical adoption of agents in daily operations

When to Use

• The task has multiple steps and repeatable judgment criteria
• Source quality and output format matter
• Human review can improve trust without blocking all progress

When Not to Use

• The task is a simple one-off answer
• The source material cannot be accessed safely
• The business cannot define what a good output looks like

Related Systems