← Research library

Field notes, July 2026: evaluation becomes a team sport

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 in July is a story about who decides what good looks like. As agents take on more kinds of work, the questions we keep returning to are less about how to test them and more about whose judgment the tests should hold.

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, and we revisit them as evidence accumulates.

Evaluation spreads beyond the builders

Earlier this year, we treated evaluation of marketing agents as the concern of whoever built them. The theme we are tracking now is a shift toward shared responsibility. We expect content leads to write rubrics and brand teams to contribute examples to voice specifications. Sales teams may rate the usefulness of account briefs, and legal reviewers may add cases to test sets.

We think this is a healthy direction. The people who know what good looks like in each area are the right people to define it. A builder can tell whether an agent followed its instructions, but a brand lead can tell whether the result sounds like the brand.

The mechanism is simple. An agent can only improve toward the standard it is measured against. If that standard reflects only the technical view, the agent gets better at being correct and no better at being useful. When the standard includes editorial, commercial, and legal judgment, improvement points in the direction the business needs.

What a shared evaluation practice contains

Shared evaluation needs structure, or it turns into a stream of opinions. The practice we would test has a few parts, each with a clear owner. We list them so teams can check what they already have.

  • A saved test set: real examples of the work, including messy and borderline cases, kept fixed between versions.
  • A rubric per agent type: short criteria that say what good looks like, written by the people who judge the work.
  • Contrasting examples: pairs of on-voice and off-voice outputs that show the line rather than describe it.
  • A review cadence: a regular time when contributors score a sample, so evaluation does not wait for a complaint.
  • A change log: a record of which instruction changed, why, and what the test set showed afterward.

The test set does the most work here. Our post on golden datasets covers how to build one that represents real work. For voice, the contrasting examples are what make a style guide usable by an agent, an idea we developed in brand voice as a specification.

Consider a hypothetical team with a test set of 40 cases for its content agent. If brand, sales, and legal each add 10 cases from their own experience, the set grows to 70. The new cases are likely to cover failure modes the builders never thought to test, which is the point.

Quiet weeks are a chance to pay down debt

The second theme is how teams use the slower weeks of summer. We expect the strongest programs to spend them on cleanup rather than new agents: trimming permissions, rewriting instructions, building test sets, and retiring agents nobody uses.

Agent debt builds quietly. An instruction gets a patch for one edge case, then another, until no one is sure which lines still matter. A tool permission granted for a one-off task stays in place for months. We described a method for clearing this in using summer's slower weeks to pay down agent debt.

Our hypothesis is that teams that do this work enter the fall with simpler systems that are easier to evaluate. The link to the first theme is direct: a shared evaluation practice is much easier to start on an agent with clean instructions and a short permission list.

Paid media and lifecycle draw attention next

The third theme is where agents go after research and content. We expect paid media and lifecycle marketing to draw the most interest. Both have direct effects on customers and budgets, which raises the stakes of a mistake.

We think the right starting point in both areas is analysis and recommendation rather than direct control. An agent that proposes a budget shift for a person to approve is far easier to evaluate than one that moves money on its own. Our framework for agentic paid media treats creative, bidding, and budget as separate loops, each with its own level of autonomy.

This is also where shared evaluation gets tested hardest. A paid media recommendation is judged by finance as well as marketing, and a lifecycle message is judged by the customer as well as the brand team. We expect rubrics in these areas to need more contributors than content rubrics do.

Year-end questions arrive early

The fourth theme is timing. We expect questions about how AI assistants will shape product research at year-end to arrive well before the season starts. Product data quality, once an operational concern, looks set to become a marketing conversation. For software and B2B teams, the parallel question is how assistants describe their products when buyers ask for comparisons.

We think the work that matters is unglamorous. Accurate product facts, consistent naming, and structured data give an assistant something reliable to draw from. Our post on product feeds for shopping agents explains why feeds now carry more weight than they used to.

Questions we keep returning to

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

  • Rubrics by agent type: how should rubrics differ for research, content, and decision-making agents?
  • Sampling by autonomy: what review sampling rates fit agents as they move up the autonomy ladder?
  • Stale test sets: how can teams detect when a test set has stopped representing real work?
  • Disagreement: when brand and sales score the same output differently, whose score should count?

Our current leaning on the last question is to keep both scores and treat disagreement as a signal. A case where two teams disagree is often a case where the brief was unclear. That is a starting view, not a finding.

What we are watching next month

August is when planning for next year begins. These are the signals we plan to watch.

  • 2027 budgets: how agent programs are funded, and whether evaluation work gets its own line.
  • Late-summer campaigns: how agents are used in back-to-school and late-summer campaigns.
  • Agent payments: how shopping agents and agent payment approaches are discussed in commerce planning.

The open question we carry into August is whether shared evaluation can scale without slowing teams down. If contributors treat scoring as part of their regular work, we expect agents to improve faster. If it feels like extra work, we expect participation to fade by the fall.

How to apply this: pick one agent and invite a reviewer from outside the team that built it to add five cases to its test set and score a sample. Keep their rubric alongside the builder's, and compare the two.

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.