← Research library

Field notes, June 2026: halfway through the year of orchestration

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Agentic marketing at the midpoint of the year is less about individual agents and more about how they work together. The questions we keep returning to are about coordination, measurement, and governance, the parts of a program that sit between the agents rather than inside them.

These are our monthly notes: themes we are tracking, questions we are studying, and what we expect. They are working hypotheses from our research and prototypes, not measurements of the market. Because this is the halfway point, we take a slightly longer view than usual.

From agents to systems

At the start of the year, the conversation we followed centered on single agents: a content agent, a research agent, a reporting agent. Each one was judged on its own output. The theme we are tracking now is what happens once a program has several of them.

Our working view is that the hard problems live between agents.

Shared context goes stale in one place and not another. A handoff drops a detail that the next agent needed. Two agents act on the same record, and no one can say which acted first. These are coordination failures, and improving any single agent does little to fix them.

Consider a hypothetical program with four agents: research, drafting, editing, and scheduling. If each passes work to the next, there are three handoffs, and each one is a place where context can be lost. Add a fifth agent that reads from the same knowledge base, and a single wrong fact now reaches every output at once. The number of agents grows by one, while the number of places a problem can travel grows faster.

This is why we have spent much of our research time on orchestration and on a coordination layer we call the hub. We describe its structure in our Agentic Hub architecture diagram. The core idea is that planning, routing, shared facts, and approvals live in one place, while each specialist agent keeps its own craft.

Measured value depends on early design

Mid-year is when programs are asked to show what the first half produced. The theme we expect to see is a wide spread between programs that can show measured value and programs that can only describe it. We think the dividing line is rarely the quality of the agents.

The mechanism is about baselines. A team that measured how long a task took before the agent arrived can compare. A team that started the agent first and asked about value later has no clean starting point, and memory is a poor substitute. Once the old process is gone, the baseline usually goes with it.

We expect a third group to face the hardest conversations: programs that expanded agent use without any measurement plan. They may have real gains, but they cannot separate those gains from noise or from changes made at the same time. A pattern worth testing is to run a short measured comparison now, on one workflow, rather than trying to reconstruct the past. Our mid-year review scorecard offers one structure for this.

Governance is catching up

The third theme is governance. We are watching for programs that write down autonomy levels, move key rules into code, and build approval workflows that can handle volume. These practices are less visible than new agents, and we expect them to matter more over the second half.

Our hypothesis is that governance often arrives after a near miss. A message almost goes out with a claim the product cannot support. A budget change runs that nobody approved. After a moment like that, rules get written quickly and sometimes too broadly.

We think the better path is to adopt these practices before the near miss. Encoding a brand or legal rule as a check that runs on every output is cheaper than adding a manual review step after the fact. It also keeps the rule consistent across agents. We describe the approach in guardrails as code, and where a rule touches advertising or privacy law, we recommend confirming its wording with counsel.

AI search becomes a planning input

The fourth theme is AI search. We expect answer engines to become part of general marketing planning, not only of SEO discussions. The change we are watching for is whether teams start to plan content around the questions buyers ask rather than the keywords they type.

Measurement here is still immature. AI answers vary by user, session, and phrasing, so a single check tells you little. We think the useful unit is a fixed set of buyer questions, checked on a regular schedule, with results tracked over time. Our post on citations, not rankings explains why being cited in an answer may matter more than position on a results page.

Questions we keep returning to

Several open questions came up repeatedly in our research this half. We do not have settled answers yet, and we list them so readers can test them alongside us.

  • One agent or several: when does a multi-agent design outperform a single well-designed agent, and when does it only add handoffs?
  • Sampling by autonomy: what is the right review sampling rate for agents at different autonomy levels?
  • Living test sets: how can teams keep test sets representative as products, offers, and audiences change?
  • Ownership: who should own the shared context that every agent in a hub reads from?

Our current leaning on the first question is to start with one agent and split only when a specific step has a clear, separate measure of quality. A split that cannot be measured separately is hard to justify. That is a starting rule, not a finding.

What we are watching next month

July begins the second half, and for many teams it brings quieter weeks. These are the signals we plan to watch.

  • Summer maintenance: whether teams use slower weeks to simplify their agent programs rather than add to them.
  • Late-summer campaigns: how agents are used in back-to-school and late-summer retail campaigns.
  • Evaluation spread: whether evaluation practices move from content agents to research, reporting, and decision-making agents.

The open question we carry into the second half is whether coordination will become a product teams buy or a discipline they build. If it is bought, we expect faster setup and less control over how context and approvals work. If it is built, we expect slower progress and a deeper understanding of where the system can fail.

How to apply this: map every handoff between your agents and note what context crosses each one. Pick one workflow, measure it against a baseline this quarter, and write down its autonomy level and approval path.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.