When to graduate to a custom agent

Advanced · ~10 min · Prerequisites: None

What you'll do

Understand the signals that tell you a recurring workflow has outgrown the assistant, and learn what a custom agent adds so you can make the right call.

Signs you've outgrown the assistant

Your assistant is great for open-ended, ad-hoc work. But some workflows repeat often enough — or need enough precision — that a dedicated agent makes more sense. Watch for these patterns:

• A recurring role. You find yourself starting threads with the same setup paragraph every time: the same role, the same tone, the same constraints. That setup belongs in a system prompt, not in every conversation.
• Fixed instructions. The workflow has rules that never change ("always cc legal on contract emails"; "never suggest deadlines shorter than two weeks"). The assistant has to be reminded; an agent is configured once.
• Its own tools and memory. The job always needs the same connector (e.g., always queries the CRM), always reads the same knowledge source, or needs a working-memory store that outlasts a single thread.
• It runs unattended. You want it to trigger on an event or run on a schedule without you initiating the thread.

If two or more of these apply, you're ready to graduate.

What an agent adds

A custom agent — built in AgentForce — gives you:

• An identity. A name, avatar, and system prompt that apply to every run without any per-thread setup.
• Scoped tools. You choose exactly which connectors, skills, and knowledge sources it can access. It can't reach things it shouldn't.
• Scheduling and triggers. The agent can run on a cron schedule or fire when a specific event happens — no human needed to start it.
• Its own memory. Persistent working memory scoped to the agent, separate from your personal assistant memory.

The agent still uses the same underlying model and capability set. The difference is configuration and scope.

Hand off the work

When you're ready to build:

1. Note the system prompt you've been reusing in your assistant threads — that becomes the agent's base instructions.
2. List the tools it always needs — those are the connectors and skills to wire in.
3. Identify where results should land — inbox notification, a doc, a Slack message.
4. Open AgentForce → New agent and configure it from there.

Your assistant threads don't disappear — you keep using the assistant for everything else. The custom agent handles only the narrowly-scoped recurring job.

Keep using both

Custom agents and your assistant are complementary. Use the assistant for exploratory, conversational work — drafting, researching, ad-hoc tasks. Use agents for reliable, repeatable, automated roles. Most power users end up with a handful of agents for their core workflows and the assistant for everything in between.

Try this next