How to use AI agents that understand your content workflow instead of starting from a blank prompt

Give your team an extra set of hands that works from the project context

When your team keeps repeating the same drafting, research, or handoff work, an AI agent can take the first pass inside the project where the context already lives. An agent is an AI teammate in your organization. It has a name, a profile picture, an AI model behind it, and a specific set of tools it is allowed to use.

Agents also do the work inside automations. Every automation action is an agent run, so the agent you build here is the one drafting your posts at three in the morning.

---

Why use agents?

Common problems solved:

• Drafting the same kind of caption over and over
• Ideas stalling because nobody has time to write the first version
• Context lost when you paste a project into a separate AI tool
• No record of what was asked or what came back

What you get:

• An assistant that already sees the project it is commenting on
• Repeatable output, because the instructions live on the agent rather than in someone’s head
• Replies posted as comments, so the whole team sees them
• Tight control over what the agent is allowed to touch

---

Before you start

You need permission to edit organization settings. If you cannot see the Agents screen, ask an owner or admin. See Managing Your Organization.

---

Step 1: Create an agent

1. Open Organization settings from the organization switcher.
2. Choose Agents.
3. Select Add agent.
4. Give it a name. This becomes its handle, so people can mention it.
5. Upload a profile image so it is easy to spot in a comment thread.

---

Step 2: Choose a model

Pick the AI model the agent runs on. The list covers the current tool-capable models, including the GPT-5 family, the o-series reasoning models, and Gemini. Faster models suit high-volume drafting. Larger models suit research and long-form outlines.

You can change the model later without rebuilding the agent.

---

Step 3: Write the system prompt

The system prompt is the agent’s standing brief. It applies to every run, so put the things that are always true here:

• Who the brand is and who it talks to
• Tone rules, including words to avoid
• Formatting habits, like caption length or where links go
• What the agent should do when it is unsure

Leave the one-off requests out. Those belong in the comment you write or the instruction on an automation action.

---

Step 4: Give it skills

Skills describe what the agent is for. An agent can carry more than one:

Skill	Suits
Producer	Keeping a project moving, summarising status
Researcher	Gathering background for a topic
Scriptwriter	Outlines, scripts, and long-form structure
Thumb Designer	Thumbnail concepts and packaging notes
Videographer	Shot lists and production notes

---

Step 5: Choose its tools

Tools are what the agent may actually do. Four exist today:

Tool	What it allows
Create Post	Draft a post on a channel and post type
Channel & Product Lookup	See which channels and post types the organization has
Project Overview	Read the project it is working in
Hook Generator	Produce hook options for a piece of content

Give an agent the smallest set that covers its job. An agent with no tools can still talk, which is often all a research or scriptwriting agent needs.

---

Step 6: Use it

Mention it in a comment

Type @ and the agent’s name in any comment on a project, page, or media file. It reads the thread and the thing being discussed, then replies as a comment. Everyone watching the thread gets notified the same way they would for a human reply.

Agents only respond when mentioned by name. They do not join conversations on their own.

Run it from an automation

Pick the agent on an automation action, add the instruction for that specific job, and choose which of the agent’s tools that action may use. See How Automations Work.

---

Tips and best practices

Prompting

• Put durable rules in the system prompt and one-off asks in the comment. Rewriting the system prompt for a single task makes every future run worse.
• Tell the agent what good looks like. A short example beats three paragraphs of description.
• Agents are told to do what was asked and nothing more, so ask for the extra passes you want.

Structuring agents

• Build one agent per job rather than one that does everything. A scriptwriter and a caption drafter want different prompts and different tools.
• Name them for the role, like Caption drafter or Research assistant, so mentions read naturally in a thread.
• Review an agent’s output for a week before pointing an automation at it.

Safety

• Only Create Post changes anything in the workspace. The other tools read.
• An agent can only see the organization it belongs to.

---

Frequently asked questions

---

Related articles

• How Automations Work: run agents on events
• Automation Recipes: setups worth copying
• Media Review and Notifications: where comment mentions usually happen