Create an AI agent
Create a Frontline AI agent with a clear role, instructions, connected context, workflow entry points, channel readiness, and a test plan before publishing.
Verify the agent role, configuration entry point, and whether it is ready to connect to tools or workflows.
How an agent is organized.
Read Overview, Settings, Flows, Conversations, Analytics, and Channels as one agent system: identity, behavior, orchestration, evidence, and publishing.
Verify the agent role, configuration entry point, and whether it is ready to connect to tools or workflows.
Verify the agent role, configuration entry point, and whether it is ready to connect to tools or workflows.
Summary
Create a Frontline AI agent with a clear role, instructions, connected context, workflow entry points, channel readiness, and a test plan before publishing.
Concepts covered
Step breakdown
- Open Agent BuilderReview the existing agents so the new role does not duplicate an existing operating job.
- Create the agent shellUse Create Agent and choose a role-based name that matches the job the agent owns.
- Configure behaviorAdd instructions, context, resources, tools, and handoff rules that keep the agent focused.
- Test before publishingUse sample conversations, edge cases, and escalation checks before connecting broader channel traffic.
What you will build
You will create the shell of a new AI teammate in Agent Builder and define the decisions needed before it can operate safely: role, instructions, context, tools, channels, workflows, testing, and publishing readiness.
The goal is not only to create an agent record; it is to make the next configuration step obvious and reviewable.
When to use it
Create an agent when a repeated operation needs AI judgment, structured extraction, conversational replies, triage, or escalation.
Use separate agents for materially different jobs, such as Sales Qualification Agent, Support Triage Agent, or HR Screening Agent, so instructions and metrics stay focused.
Before you start
Write the job the agent owns, the channels where it may respond, the CRM or table context it can use, the handoff rule, and the examples you will test.
Prepare sample prompts and sample conversations that contain no customer data, credentials, private URLs, or production identifiers.
Step-by-step implementation
Open Studio, then Agent Builder, and review the existing agent list before creating a new one.
Select Create Agent and give the agent a role-based name, such as Support Triage Agent or Sales Qualification Agent.
Open the new agent and review identity, owner, model or provider setting, temperature, instructions, knowledge, tools, playbooks, deployment, and chat customization.
Add concise instructions that state the job, allowed context, response style, escalation rule, and what the agent must never do.
Attach only the resources, tools, workflows, channels, and tables the agent needs for this job.
Run tests with sample conversations, missing context, edge cases, and escalation scenarios before expanding scope.
How to verify it worked
The agent should have a clear role, focused instructions, connected context, and a documented path into a workflow or channel before it is treated as operational.
Review chat tests, workflow logs, analytics, escalations, and Max Activity to confirm the agent supports the intended outcome.
Common mistakes
Do not create one broad agent for unrelated jobs.
Do not attach every resource by default; noisy context makes answers harder to verify.
Do not publish before the agent has passed sample prompts, missing-context tests, and handoff tests.
Troubleshooting
If answers are generic, tighten the instructions and pass more precise CRM, table, or workflow variables.
If the agent takes the wrong action, check which tools and workflows are attached and whether the instructions define when to use them.
If escalation fails, verify the workflow path, assignment, channel, and Max Activity step.
Clear next step
After the agent shell is created, configure its instructions and connect it to one workflow with a sample test path before adding more channels.
Implementation path
1. Open Studio, then Agent Builder.
2. Review the existing Sales, HR, Support, and Marketing agents so you understand what a production agent looks like.
3. Click Create Agent and name the new role by job, not technology.
4. Choose the model/provider for the job. Frontline is model-agnostic: Claude, GPT, Gemini, DeepSeek, and other supported providers can be selected as they are configured.
5. Add instructions that define role, tone, allowed actions, escalation rules, and output format.
6. Attach knowledge bases, tools, playbooks, tables, and resources.
7. Connect channels only where the agent should communicate.
8. Connect flows and CRM context before publishing.
9. Test with realistic conversations and review Analytics before expanding scope.
How real agents connect to operations
Sales Agent qualifies leads, uses CRM context, triggers lead capture flows, and hands qualified opportunities to Sales.
Support Agent reads customer messages, checks ticket context, drafts safe replies, and escalates risky cases.
HR Agent reviews candidate or employee context against approved criteria and prepares People-team next steps.
Marketing Agent turns campaign or social signals into captured leads, segments, enrichment, and follow-up workflows.
What to validate before launch
Validate selected model/provider, instructions, temperature, reasoning iterations, connected knowledge, tools, playbooks, flows, channels, and CRM permissions.
Then review Conversations and Analytics. If the agent cannot explain or repeat the expected operating behavior, it is not ready for customer-facing work.
Transcript
Open searchable transcript
FAQs
Should one agent handle every department?
No. Create focused agents for different operating jobs so instructions, context, tests, and analytics stay understandable.
What should I test first?
Test normal sample conversations, missing context, ambiguous intent, escalation, and the expected workflow output.
When is an agent ready for a channel?
When its instructions, connected context, handoff rule, logs, analytics, and sample responses have been reviewed.
What is a Frontline agent?
A Frontline agent is an operational AI teammate with identity, behavior, prompts, connected context, channels, flows, conversations, analytics, and permissions.
When should I create a new agent instead of reusing an existing one?
Create a new agent when the operational role, tool access, escalation rules, or expected output is meaningfully different. Reuse an agent when the same job simply needs another workflow entry point.
How should prompts connect to workflows?
Prompts should describe the agent's job and output in a way the next workflow node can use. If a workflow branches on the result, ask the agent for structured values or a clear decision.
What memory should an agent have?
Give the agent the minimum useful memory: approved resources, relevant CRM context, table data, and current workflow state. Too much memory makes behavior harder to test and audit.
How do I test an agent before production?
Test the agent with realistic conversations, missing context, edge cases, escalation scenarios, and expected workflow outputs. Review tone, accuracy, tool use, and handoff behavior.
What permissions should an agent receive?
Give agents only the integrations and actions required for their role. If the agent only drafts or classifies, it may not need write access to external systems.
How do agents connect to channels?
Channels define where an agent can communicate. Use WhatsApp, Instagram, Messenger, or other channels with explicit routing, approved templates, and handoff rules.
How should teams monitor an agent after launch?
Review conversations, analytics, workflow logs, escalation quality, and Max Activity. Monitoring should show both AI quality and the operational outcome the agent supports.