Chatting with agents

Each AgentForce agent has its own chat surface at /agentforce/<slug>/chat. The interface is similar to your personal assistant — composer at the bottom, message stream in the middle, thread sidebar on the side — but every conversation is scoped to the agent you're talking to.

Threads

Each agent maintains its own list of conversations. Open the agent and you'll either land on a fresh thread (auto-created on first visit) or the most recent one.

Creating a new thread: Click + New conversation in the thread sidebar. A fresh thread opens with a unique URL like /agentforce/sales/chat?thread=.... You can have as many threads as you like — they don't expire.

Switching threads: Click any thread in the sidebar. The URL updates and the message history loads.

Thread titles: New threads start untitled. Once you exchange a few messages, the agent auto-generates a short title based on the topic. Titles can be edited later if you want.

Persistence: Threads are stored in Mastra's memory — refresh, sign out, and come back days later, and the same threads are still there with full message history.

Suggested starter prompts

Every agent's empty state shows three suggested starter prompts tailored to its specialty. Click any one to drop it into the composer (the prompt isn't sent yet — you can edit it before submitting).

The suggestions are a quick way to discover what the agent is good at. They're worded as if you'd written them yourself, so they make sense as conversation openers.

If none of the suggestions match what you need, just type your own.

File attachments

The composer accepts file attachments — drag onto the chat or click the attach button. Supported formats include text, PDFs, and common image types. Once attached, the agent can reference the file in its response.

Use cases:

• Support agent: drop a customer's email screenshot to draft a tone-matched reply
• Research agent: drop a PDF report to synthesize the key findings
• Content agent: drop a draft blog post to edit
• Sales agent: drop a prospect's LinkedIn export to summarize their activity

The file goes into the agent's context for the current message; the agent decides how to use it.

Mid-conversation tool calls

As the agent works, you'll see small tool-call pills appear in the conversation. Common ones:

• search_kb — searched the knowledge base
• search_org_contacts — looked up a person in the org's CRM (sales)
• read_file — opened an attached file

The pills are useful as a sanity check — they tell you what the agent did to produce its answer. If you see no tool calls and the agent claims to have looked something up, it didn't actually do so.

Tips for good agent conversations

• Start specific. "Draft an apology email to a customer whose shipment is 5 days late" lands better than "Help with a customer issue".
• Lean on the agent's specialty. The Support agent is tuned for empathy; ask it about messaging tone, not about server architecture.
• Iterate within the thread. Each agent has its own personality and remembers what you've discussed; threading multiple turns gets you closer than starting over.
• Don't switch agents mid-thread. If you realize the wrong agent is handling the work, copy the relevant content into the right agent's thread and start fresh.

Switching back to your personal assistant

AgentForce agents are great for specific roles, but your personal assistant is the one that knows you, your memory, and your full workspace context. Use the agents for specialized work and the assistant for everything else.