Build a marketing coworker

In this tutorial you build an agent that works on your marketing, teach it your company and your voice, correct it once, and then hand the recurring part of its job to a schedule. By the end, the agent produces next week's content plan every Monday morning without you opening a chat.

Before you start

This tutorial assumes:

• You finished Build your first agent, so you know the project home, the playground, and how an agent configures itself while you talk to it.
• Your project has a model provider key. If the chat input is read-only and asks you to connect a model, add a key under Settings > LLMs first.
• You have about thirty minutes. The example company below is fictional. Replace its facts with yours as you go, or copy it as written and swap them later.

Times in the scheduling dialog are UTC.

Step 1. Describe the coworker

Open Sidebar > Home. Under What do you want to build?, type the description of the agent you want and press Create agent.

Use this description:

A marketing coworker for a small B2B software company. It drafts and reviews
marketing copy, keeps our product facts straight, and plans what we publish each
week. It should work from a written brief about the company rather than guessing.

Agenta creates the agent, names it, and opens its playground. The conversation opens with a card summarising what it is about to set up and a Start → button.

Leave Playground mode on Build. The left column is Configuration and the right column is the conversation, so you can watch the configuration change while you talk.

The agent it created works, and it knows nothing about you. That is the next step.

Step 2. Give it your context

An agent's instructions are its identity. They live in a file called AGENTS.md and they are sent at the start of every session, so whatever is in them is guaranteed to be there every time the agent thinks. See Agents for the full picture.

In the Configuration column, open the Instructions section and click the AGENTS.md row. The editor opens with an Edit and Preview control at the top. Click Source in the toolbar to edit the raw markdown.

Replace what is there with this, then press Save:

# Role

You are the marketing coworker for Kettleworks. You draft copy, review copy I
write, and help me decide what to publish next.

# The company

Kettleworks sells inventory software to independent coffee roasters. It tracks
green coffee, roast batches, and wholesale orders in one place. We are six
people. The product is self-serve at $79 a month and there are no sales calls.

# Who we write for

Owner-operators who roast in the morning and do the books at night. They are not
software buyers. Most of them are moving off a spreadsheet, and that spreadsheet
is what we get compared against.

# How to write

- Short sentences. No adjective a competitor could also use about themselves.
- Open with the concrete thing the reader gets, not with the problem.
- Say "roaster", not "customer" or "user".
- No exclamation marks.

# How to talk to me

- Answer first, explain after. Do not restate my question back to me.
- When you draft something, give me one version, not three, unless I ask for
  options.

Four sections, all facts, no filler. Everything in that file is resent on every turn, which is why it holds only what has to be present every time. Reference material the agent can look up when it needs it goes in a file instead, and you add that in the next step.

You can also ask the agent to make this edit for you in the chat. Writing it yourself here means you see exactly what the agent will read.

Step 3. Work with it, then fold the correction back in

Ask it for something real. In the conversation on the right, send:

Draft the announcement email for batch scheduling. It lets a roaster plan a
week of roasts in advance and see green coffee shortfalls before they happen.

Read the draft. The tone should already be close, because you wrote the tone rules down. What it cannot get right yet is anything factual: it does not know your prices, which integrations exist, or which features have shipped, so if the draft needs one it will make it up.

Say the draft claims a Shopify integration and "over a thousand roasters". Both are wrong. There are two ways to answer that, and this tutorial uses both.

Put the facts in a file. Send this in the conversation:

Write our product facts to agent-files/product-facts.md:

Pricing: $79 a month, one plan, billed monthly. 14-day trial, no card.
Integrations that exist: Cropster import, QuickBooks Online export, CSV import
and export.
Not built, never mention: Shopify, Square, WooCommerce, mobile app,
multi-roastery accounts.
Public numbers: 400+ roasters on the product; first roast batch logged in
under 10 minutes.

agent-files/ is the agent folder. Every session of this agent sees it and it survives after a chat ends, so a fact you write there today is still there next month. The other folder in the file browser belongs to the current chat and disappears with it. See Files and knowledge.

The agent stops before it writes and shows a card titled Approval needed to continue. By default an agent reads on its own and asks before it writes. Press Approve.

The card also offers to always allow that one tool for this agent. Its note reads Applies when you approve; commit to use it in triggers. Leave it alone for now. Step 4 handles permissions in one place.

Point at the file from the instructions. A file only helps if the agent knows to open it. Open AGENTS.md again and add one line to # How to write:

 # How to write

 - Short sentences. No adjective a competitor could also use about themselves.
 - Open with the concrete thing the reader gets, not with the problem.
 - Say "roaster", not "customer" or "user".
 - No exclamation marks.
+- Every product fact (price, integration, number, shipped feature) comes from
+  `agent-files/product-facts.md`. Read it before you draft. If a fact is not in
+  that file, ask me instead of writing it.

Save, then send Redo that email in the conversation. The agent reads the file first and the invented Shopify line is gone.

That is the loop for the rest of this agent's life. One line of instructions per correction, with the detail behind it in a file. The instructions stay short and turn into a map of where things are. Write your agent's instructions covers the writing, and Build a knowledge base for your agent covers the files.

Step 4. Hand the recurring part to a schedule

Some of this work repeats on a clock. Planning what to publish next week is a Monday job, and there is no reason for you to be in the chat while it happens. An agent that runs on a schedule is an automation: same instructions, same files, same tools, different starter.

Three things change when nobody is watching.

• The agent cannot ask you anything. A question mid-run has no one to answer it, so the answer has to be in the instructions before the run starts.
• Permissions have to be settled in advance. A run that stops for approval waits for an approval that is not coming.
• The output has to land somewhere you will look. Here that is a file in the agent folder.

4a. Write the job down

Open AGENTS.md and add this section at the end, then press Save:

# Weekly content plan

Every Monday you produce next week's content plan without me.

- Read `agent-files/product-facts.md`, then the four most recent plans in
  `agent-files/plans/`.
- Propose five items: two blog posts, two emails, one changelog note. For each
  one give a title, the roaster problem it answers, and two sentences of angle.
- Do not repeat anything that appears in the last four plans.
- Write the result to `agent-files/plans/<YYYY-MM-DD>-content-plan.md`, where the
  date is the Monday of the coming week.
- If something you need is missing, do not ask me. Write the plan anyway and list
  what was missing under a final heading called "Gaps".

The last two bullets are the ones that matter for an unattended run. One says where the output goes, so you can find it. The other tells the agent what to do instead of asking, so a missing fact ends up as a note in the file rather than as a run that never finishes.

4b. Settle the permissions

In Configuration, open the Advanced section and go to Permissions. Set Policy to Allow all, then press Save.

This agent's tools read and write files in its own folders and nothing else, so Allow all lets the Monday run finish on its own. The Auto-approve field in the same section lists any single tool you granted from an approval card, which is the narrower way to reach the same result.

warning

Allow all means every tool call runs without asking. Check the tool list before you set it on an agent that can send messages or change records in a connected app. See Control what an agent can do.

4c. Commit the version

A schedule runs one exact version of the configuration, so the work from the last three steps has to be committed before a schedule can point at it.

At the top of the Configuration column, press Commit. The left pane lists what changed. On the right, keep New version, edit the commit message if you want, and press Commit.

Every commit is kept, so you can see what the agent looked like at any point and compare. See Sessions and versions.

4d. Create the schedule

In Configuration, scroll to the Triggers heading and its Schedules section. Press Add schedule. The New schedule dialog has four parts.

1. Name. Type Weekly content plan in Schedule name.

2. When should it run? Choose Weekly. Under On these days pick Mon. Under At these times (UTC) set 09:00. A green line confirms the schedule it parsed and the next time it fires.

3. Which version runs? Leave it on Pinned and pick the variant you just committed. Pinned runs one exact variant and revision, so a later draft in the playground cannot change what Monday does.

4. What should the agent do? This text is sent to the agent as the user message on every run. Write:

   Produce next week's content plan, following the "Weekly content plan" section
   of your instructions.

The message is short because the instructions carry the job. That is the trade: work you would have explained in a chat has to be written down once, in a place the agent reads every time.

4e. Watch it run once, then create it

You do not have to wait until Monday to see the result. Before you save the dialog, press Run in playground in its footer. That runs the message once in the playground, so you can watch the whole job end to end.

Once the run has finished, leave Active on and press Create. The schedule is live from the next Monday at 09:00 UTC.

Now go and find the output where you told the agent to put it. Switch Playground mode to Chat, open the files panel on the right, and press Open the files drawer. Browse into agent-files/ and open plans/.

The plan is there, written by a run you did not type into. Every Monday at 09:00 UTC another one appears next to it, and each run reads the four before it.

What you built

• An agent that knows your company, your audience, and how you want it to write, because you put that in AGENTS.md.
• A fact file in the agent folder that the instructions point at, so corrections do not make the instructions longer.
• A committed version, so this configuration has a history.
• A weekly run that starts on a clock, finishes on its own, and leaves its output in a folder you can open.

Next

• Trigger an automation from an app starts a run when something happens in a connected app instead of on a clock.
• Tools and integrations covers giving the agent actions in other systems, including sending the plan somewhere instead of writing it to a file.
• Skills covers packaging a repeated procedure so the agent loads it only when it is relevant.
• Cost and usage covers what these runs cost.