How to buildAI marketing agents.

Do you need an agent or a workflow?

Anthropic distinguishes workflows, where predefined code paths coordinate models and tools, from agents, where the model dynamically directs parts of the process. That distinction is useful for marketing: a fixed report-generation task and an open-ended investigation do not need the same amount of autonomy.

Begin with the simplest structure that can complete the task. A campaign summary might only need a data query, a calculation step, and a model that explains the result. Letting a model choose arbitrary tools adds little value if the sequence is already known.

Reference: Anthropic — Building effective agents

Choose a task with a reviewable output

Our suggested first project is a draft performance brief. It is useful, bounded, and easy for a person to inspect. The system receives an account and reporting period, retrieves permitted campaign data, calculates a few agreed metrics, and produces a draft with references to the underlying records.

The completion criteria should be explicit: correct dates, consistent currency, no invented metrics, a clear distinction between observed data and proposed explanations, and a link back to the records used. “A persuasive report” is a much weaker specification than “a report whose numbers can be reconciled.”

Give each tool a small, clear contract

Treat a tool as a controlled interface rather than a convenient escape hatch. A reporting tool can accept a date range and a permitted account ID. A draft-creation tool can store a proposal. Neither needs permission to change a campaign budget.

Illustrative workflow: fetch campaign totals, calculate changes using ordinary code, ask the model to explain relevant differences, then save a draft for review. Keep source data separate from the model’s interpretation. If a data request fails, return an explicit error rather than asking the model to fill the gap.

Separate a recommendation from an action

Reading performance data, writing a draft, and changing live ad spend are different capabilities. In a proposed implementation, keep the last capability behind a separate approval step. Present the account, intended change, relevant evidence, and scope of the authorization before execution.

Handle retries deliberately. A repeated request should not create duplicate campaigns or apply a budget change twice. Keep a record of the action request and its outcome so a reviewer can tell whether a failed response means the action failed or merely that its acknowledgement was lost.

Evaluate the whole task

A model’s fluent answer is only one part of the system. Test missing rows, incompatible currencies, empty date ranges, duplicate requests, and unavailable tools. Include cases where doing nothing is the correct outcome. Review whether the final artifact is accurate and whether every action stayed within its allowed scope.

Atakan AI is developing Ads Agent API around Atakan 1.0, a planned model and tool orchestration interface for marketing tasks. The workflow on our site is conceptual. APIs, supported tools, and developer access are still in development; the waitlist is the place to register interest in future updates.