Celeste Custom Agents: Your Team’s Most Productive Analyst Never Clocks Out

Most AI products in marketing are built around the same interaction model. You ask a question and act on the answer. The loop has real value but it also has a ceiling.

The ceiling shows up in the work that happens on a schedule: weekly pacing reviews, cross-channel performance summaries, creative attribute analysis that needs to run every Monday before someone can make a budget call. These are not one-off questions. They are recurring workflows, and every week someone on your team has to start them from scratch.

Assistance is reactive. Agency is not.

When we built Celeste, we started with the conversational layer because that is where the most immediate value was. Marketers could ask questions in plain language, get answers grounded in their actual campaign data, and act on them faster than any reporting workflow allowed. More than 1,500 users across 200+ organizations are doing exactly that today, generating tens of thousands of interactions every month.

But the teams getting the most out of Celeste told us something else: they kept running the same prompts, on the same schedule, week after week. The prompt was not the problem. The repetition was.

Custom Agents solve that. An agent is a Celeste analyst you set up once. You define its focus, write its instructions, and give it a schedule, and from there it runs on its own. Nobody has to remember to start it, and the analysis lands in your inbox on the cadence you chose.

Picture a team running weekly pacing reviews across five publishers. Instead of asking Celeste for a summary every Monday, they build a pacing agent and set it to run Sunday night. The report is waiting before the week starts, in the same format every time, and the analyst who used to spend an hour building it has that hour back.

What makes these agents reliable

A lot of tools can generate a report. The harder problem is generating a report you would trust enough to share with a client or act on without double-checking. 

That reliability depends on three things:

1. Data foundation

Celeste is the only AI agent operating inside a platform that unifies retail media, paid search, paid social, and app data in one place, which means agents are not working from a siloed snapshot of performance. They are working from the same cross-channel view your team relies on. Approximately 68% of Celeste users manage three or more publishers, and the agents see all of it.

2. Explainability

Every Celeste response, including agent-generated ones, includes Answer Breakdown: a full trace of how the agent reached its conclusion, which data it pulled, what it prioritized, and where the insight came from. That is not a feature for the skeptics. It is what makes the output usable in a client meeting or a leadership review.

3. Context

Agents work within the instructions, inputs, and scheduling parameters you define. You set what they focus on, decide the output format, and control whether they run on a schedule or wait for a trigger. The agent executes inside those parameters every time, which is what makes the output consistent enough to delegate on.

Reliability at this level is what allows teams to actually stop doing the work themselves, not just use the output. That is a different standard than most AI tools are designed to meet.

Who can build one

We made a deliberate decision with the template library: the people who should be able to build agents are not the people who understand prompt engineering. They are the people who understand the workflows.

Pre-built agent templates are designed by media experts and cover the recurring analysis patterns most teams run on a weekly basis, so a new Celeste user can deploy a working agent from a template in minutes. An experienced user can convert any prompt they have already written into an agent in a few clicks. Neither path requires deep configuration, and both paths produce agents grounded in the same unified data foundation.

I have seen this matter most at agencies. The bottleneck is not willingness to use AI. It is the distance between what a tool can do and what a team can actually set up, and the template library exists to shorten that distance considerably.

An agency running the same reporting workflows across 20 client accounts does not need 20 different prompts. They build one agent template, deploy it at scale, and the analysis runs consistently across every account on whatever cadence they set.

What agents can and can’t do today

Custom Agents are built for analysis and monitoring. That includes weekly pacing summaries across all publishers delivered before the week starts, performance shift detection and anomaly flagging on a schedule, cross-channel reporting in a consistent format, audience-level insights for offsite campaigns, and creative performance analysis using AI-tagged attributes tied to outcomes.

What agents cannot do yet is take action on bids, budgets, or campaigns. Autonomous optimization is on the 2026 roadmap, and the analysis and monitoring layer, the part that consumes the most time in most teams, is what is live today. We designed agents to do what they can do reliably before expanding what they can do autonomously. The teams building agent workflows now will be in the best position when the action layer follows.

The broader context

At Skai ShopAble 2026, we described the shift we are seeing in commerce media as a move toward a new operating model. 

According to Gil Sadeh, president at Skai, “A new org design with a hybrid organization of people directing intelligent agents. Reporting, optimization, and analysis happening continuously, in the flow of work.” 

Custom Agents are where that model becomes concrete for Celeste users today, and they are also the foundation of what comes next in Skai Studio, where teams will orchestrate a full AI marketing workforce executing strategy end-to-end.

The teams I have spoken to who are furthest along in this shift share a specific characteristic. They are not the ones who adopted AI most aggressively. They are the ones who were most precise about which parts of their workflows should stay human and which parts should run automatically. That precision is what separates useful automation from noise. Agents do not change the judgment calls your team needs to make. They change how much time you have to make them well.

Already a Skai customer? Contact your success team to get started with Custom Agents. New to Skai? Schedule a demo.

---

Frequently Asked Questions