Most of the campaign reporting work we hand to agents is not long-running. It is long-horizon. A weekly performance report doesn't need an agent that thinks for hours without stopping. It needs one that picks up on Tuesday where it left off on Friday, remembers why it flagged a cost spike, and knows who was supposed to look at it.
That distinction came through clearly in a recent Latent Space podcast conversation with Aman Singh, lead AI engineer at Arca, an AI-native wealth management firm. Financial advice and paid media are different businesses, but the shape of the agent problem is surprisingly close. Below we unpack what Arca described, then translate it into a design for multi-day campaign monitoring and weekly reporting agents. The design covers what to store, what to re-check and when the agent should stop and ask a person.
Long-horizon is not long-running
Latent Space reported that Singh described AI work in financial services as typically long-horizon, although not long-running, and noted that agent tasks at Arca can take up to ten days. In his words: "It's very stop-start, there's many actors in between, many decision points, many handoffs."
Replace "client portfolio" with "Google Ads account" and you have a fair description of campaign management. A new creative goes live on Monday. Nobody should judge it on Monday afternoon. Conversions keep arriving for days after the click. The client approves a budget shift by email on Wednesday. The landing page gets swapped on Thursday without anyone telling the media team. By Friday, the weekly report has to make sense of all of it.
An agent that starts from a blank context every run can't do that job well. It will either re-discover the same anomaly every morning or miss the fact that last week's anomaly has a known cause. The fix isn't a bigger context window. It is deliberate, structured memory.
What Arca's setup gets right
Three ideas from the conversation are worth borrowing directly.
Persistent memory per client, organised around open work. Singh described the questions Arca's memory system answers for every client: "What has happened so far? What are we tracking for next? What needs to happen immediately in the next couple of days [and] what are we tracking for the long term? Which individual is responsible for which action?" He added that all of this is tracked by memory systems the agents contribute to. That is a to-do list with history and owners, not a chat log. That distinction matters.
A harness they own. According to Latent Space, Arca built its own harness, which selects the model for each task based on internal benchmarks. Owning it rather than using a managed agent harness also means they can change model providers if they need to.
Deterministic safety, with humans on anything risky. The agent's role is limited to analysis and proposing solutions. Trading and other money movement stay with a human advisor. "All of our safety restrictions are fully deterministic," Singh said. "And they take the decision making off the agent's hands."
The line that sums it up, in Singh's phrasing, is that "agents prompt people more than people prompt agents." That's the right mental model for a reporting agent too. Its job is not to wait for someone to ask how the account is doing. Its job is to tell the right person, at the right moment, what needs their attention.
What a campaign agent should store
We think about memory for a campaign agent in six buckets. Each one answers a question the agent would otherwise have to guess at.
- Context changes. Tag changes, landing page swaps, consent banner updates, new conversion actions, offline imports switched on. These change what the numbers mean. They are the single most useful thing an agent can remember.
- Open threads. Every anomaly the agent has flagged, with its hypothesis, the evidence it used and a status: open, explained, resolved or dismissed by a human.
- Tracking-for-next items. Things that shouldn't be judged yet but need a scheduled look. One example is a new ad group that should be reviewed once it has enough data, with the condition written down.
- Owners. Who on the agency side and who on the client side is responsible for each open action. Arca tracks this explicitly, and so should you. A flag with no owner is just noise.
- Decisions and their reasons. Budget moved from campaign A to campaign B, approved by a named person, for a stated reason. Next month, when someone asks why, the agent can answer.
- Baseline definitions. Which date range, which conversion action, which attribution setting and which query produced the comparison figure in last week's report.
Notice what is missing: the metrics themselves. We deliberately don't treat stored numbers as truth. Memory should hold pointers, such as the query, the date range and the definition, rather than the result. Numbers get re-fetched.
What to re-check on every run
Memory is a claim about the past. Anything that can drift since the claim was written has to be verified before the agent leans on it.
Re-query the numbers. Last week's conversion count is not final if conversions are still being attributed to those clicks. The agent should rebuild comparison figures from source each run, using the stored definitions, and note when a previously reported figure has moved. If you're building this on the GA4 export, our walkthrough on conversational agents on GA4 BigQuery exports covers how to give an agent safe, structured query access to that data.
Re-verify measurement before interpreting performance. A drop in conversions is a performance story or a tracking story. The agent should check which one before it writes a sentence. That means checking whether the GTM container was published since the last run, whether the conversion action still fires, and whether the event counts in the warehouse line up with what the ad platform reports.
Re-check open threads, not just the dashboard. For each open thread the agent should ask three things. Is the condition still true? Has the owner responded? Has the window we agreed to wait for passed? This is where most of the value of persistence shows up. The agent closes loops instead of opening new ones.
One design principle helps a lot here. Keep the model that reasons about the account separate from the tools that fetch and change things. We wrote about that pattern in separating brain from hands in MCP marketing agents. It's what makes it practical to re-check aggressively without giving the reasoning layer write access to anything.
When the agent should prompt a human
Arca's rule is simple: analysis and proposals belong to the agent, and money movement belongs to a person. For campaign work we'd draw the line in the same place. Budget changes, bid strategy changes, pausing or enabling campaigns and anything else that spends or stops spending client money should be proposed by the agent and approved by a human.
The important part is where that rule lives. It belongs in the harness, as deterministic code, not in the prompt. A prompt that says "never change budgets without approval" is a request. A harness that simply has no write tool for budgets, or routes every write through an approval step, is a guarantee.
Beyond money, we'd have the agent prompt a person in four situations:
- A threshold you defined in advance is crossed. The threshold is set by people, per account, and stored in memory, not improvised by the model.
- Measurement looks broken. A suspected tracking failure is more urgent than any performance insight. Everything downstream depends on it.
- An open thread has aged out. The owner hasn't responded within the agreed window, so the agent escalates rather than re-flagging silently every day.
- Fresh data contradicts memory. If the stored explanation for an anomaly no longer fits the numbers, a human should decide what's true.
The format of the prompt matters as much as the timing. A good one states what changed, shows the evidence, proposes an action, names the owner and says what happens if nobody responds. If the person can act on it in a minute, the agent is doing its job. If they have to open three dashboards to understand it, it isn't.
Own the harness, even a small one
You don't need Arca's engineering team to apply the harness lesson. For most marketing teams, the harness is a workflow tool or a set of scripts. It decides when the agent runs, which tools it can call, where memory is read from and written to, and which actions require approval. Keeping that layer in your hands gives you two things. Safety rules are deterministic, and you can swap the model underneath without rebuilding the process.
Arca also described using human advisor feedback, together with agent traces, as input to a meta-agent that suggests changes to the harness. A lightweight version is within reach for any agency. Every time an account manager edits a draft report, dismisses a flag or overrides a proposal, record the edit and the reason. Over a few weeks that record tells you which thresholds are too twitchy, which explanations the agent keeps getting wrong and which checks it should run earlier.
Where to start
Pick one account and one weekly report. Write down the six memory buckets for it by hand, then build an agent that reads that memory, re-checks it against live data and drafts the report with its open threads and owners listed. Keep every write action behind approval. Once the draft consistently matches what your account manager would have written, add the next account.
If you want help designing the memory layer, the harness or the MCP connections between Claude and your ads, analytics and tagging stack, talk to us at Bitegrico. This is the implementation work we do every day.