Coordination has a ceiling, and adding agents does not raise it. Martech Futurist | August 4, 2026

My last post highlighted insights that identified the current constraint on marketing throughput. Teams pointed AI capacity at the fast steps, drafting and image generation, while approvals, assembly, and cross-team coordination stayed manual. Launch dates kept slipping because the slow steps never got touched.

The logical next move is to point agents at coordination itself. Research published in the past week sets a measurable limit on how far that move goes. Collective accuracy in a multi-agent system peaks at a specific headcount and falls after it. Separate findings show autonomous agents skipping assigned work and reporting completion they did not achieve. Together these define the operating problem for the next two quarters: as agent count and task length rise, verification cost rises faster than throughput. The organizations getting production results are the ones bounding what each agent may do before they add another one.

Agents widen the scope of work, and move people into supervision

Harvard Business Review | July 29, 2026Research: How AI Agents Broaden the Scope of Knowledge Work

Jeremy Yang, Kate Zyskowski, Noah Yonack, and Jerry Ma studied production data from Perplexity's conversational assistant and its autonomous agent, matching near-identical query pairs across both products over a three-month window. The agent performed 26 minutes of autonomous work per session against 33 seconds for search, cut task time and cost, and expanded the range of work people attempted, including projects that crossed occupational boundaries and multi-step work that rarely appeared with a standard assistant.

The finding that matters for marketing operations sits in the authors' conclusion. As agents absorb execution, people move from operating to supervising, and leaders have to redesign roles, workflows, and oversight to govern the autonomy they just granted. This describes exactly the mechanism that produced last week's numbers. When execution gets cheap, teams attempt more work, and the supervisory load lands on the same approval chains that were already the constraint. The study is worth reading with its provenance in view: Perplexity supplied the data and three of the four authors work there.

Multi-agent accuracy peaks near 16 agents, then declines‍ ‍

NTT Research and Harvard Center for Brain Science | July 28, 2026How Many AI Agents Are Too Many?

Hidenori Tanaka and Elizabeth Pavlova of NTT Research's Physics of Artificial Intelligence Lab, working with Harvard's Center for Brain Science, built an experiment called the Flag Game. Each agent sees one randomly assigned fragment of a hidden flag, and the group has to communicate to identify it. Success requires 85% of agents to converge on one answer. Collective accuracy peaked at roughly 16 agents. Below that count, the group gathered too little evidence to reach consensus. Above it, communication degraded and the agents polarized into competing camps.

This is the clearest quantitative answer yet to the question a CMO asks after reading a productivity study: how many agents should we run. The number itself will vary by task and architecture, and one benchmark presented at an ICML workshop does not settle enterprise design. The shape of the curve is the transferable finding. Multi-agent systems have an operating range, and performance falls outside it in both directions. Marketing organizations planning agent fleets across content, media, lifecycle, and analytics should treat headcount as a design variable with an optimum, and should measure where their own optimum sits.

It reminds me of what Chris O'Neill, CEO at GrowthLoop, said when I interviewed him on The Agile Brand podcast: "Marketing cycles are too darn slow. There's manual steps at every step of the cycle…and that really holds companies back." The manual steps are real. The research says the fix runs through workflow design before it runs through agent count.

The verification bill comes due as autonomy extends

MarketingProfs | July 31, 2026AI Update, July 31, 2026

Three items in this week's roundup describe one failure mode. Researchers and developers report autonomous systems taking shortcuts, skipping assigned work, misrepresenting tasks as complete, and optimizing for evaluation criteria over instructions, with the problems traced to training incentives that reward efficiency over correctness. Andon Labs ran frontier models through a year-long simulated vending machine business and observed collusion, deception, and broken agreements among the highest-scoring models. And an OpenAI agent that breached Hugging Face operated outside its testing environment for roughly a week before the company identified it as the source.

Each of these lands on the Trust and Verification layer in the capability framework I use, in the specific form that layer takes under Orchestration. The question is whether the agent did what it reports, and answering it requires an audit trail of actions. Reviewing the output alone leaves the question open. My framework treats the mechanism spectrum, from deterministic rules through predictive and generative to agentic, as a design variable, and the discipline is to use the lightest mechanism that clears the task's accuracy bar. Reaching further along that spectrum costs twice, once in compute and once in verification. A detection lag measured in days is what the second cost looks like in production.

Michelle Cooper, Chief Marketing Officer at NiCE, made the design point when I interviewed her on The Agile Brand podcast: "A lot of companies are kind of approaching this AI era from a technology, right, from a tool perspective. Where you really need to start is in thinking through what are the customer moments that matter, the business processes you're trying to evolve, and the outcomes you're ultimately trying to get to."

Bounded authority makes an unsupervised handoff work

Visa and Lianlian DigiTech | July 24, 2026First live B2B agentic transaction via LoopXPay

Visa and Lianlian DigiTech completed what they describe as Greater China's first live business-to-business purchase executed by an AI agent. Lianlian's LoopXPay agent identified the purchasing requirement, recommended suppliers, compared options, placed the order, and executed payment in a single workflow, operating inside pre-defined spending controls and approval parameters. LoopXPay is registered in Visa's Agentic Directory, which supports the Trusted Agent Protocol and lets businesses and merchants identify verified agents.

This is the constructive half of the week. A multi-step commercial action ran end to end with no human in the sequence, and it worked because the authority was scoped before the agent started. Every element maps to the Identity and Permissions layer: on whose behalf, with what authority, within what bounds. Compliance here shifts toward knowing the agent, which is a durable pattern for marketing as agents begin transacting in media buying and procurement. Marketing teams should read the registry and the protocol as the parts worth copying, because they are what convert an agent from a demonstration into a system an operator can hold accountable.

Key insights

Set an agent headcount, and measure the optimum. The Flag Game result establishes that more agents degrade collective accuracy past a point. Treat fleet size as a tuned parameter across each workflow, and instrument for the decline.

Budget verification alongside compute. The July 31 findings show detection lag and misreported completion as ordinary operating conditions. Cost models that count tokens and omit review time will understate the true unit cost of autonomy.

Scope authority before adding capacity. The Visa implementation ran unsupervised because spending limits, approval parameters, and agent identity existed first. That ordering is the transferable lesson.

Redesign the approval chain alongside the tooling. The HBR research shows agents pulling people into supervision and expanding the volume of work attempted. Approval capacity has to be redesigned to absorb that, or the coordination constraint from last week's edition gets worse as capacity grows.

This maps to Agile Brand Principle 3, operationalizing adaptivity, which asks organizations to make collaboration and change part of how the work runs. It also sits on the Book 2 commitment that humans stay accountable for direction even when analysis and execution are automated. Descriptive work moves to the agents. The normative call about what the organization should do stays with people.

A few closing thoughts

I have run a couple companies and sat through enough acquisition diligence to recognize the pattern in this week's research. When a team tells you throughput is the problem, the numbers usually show coordination is the problem, and coordination rarely improves by adding headcount to it. That was true of people, and this week's evidence says it is true of agents.

The twelve years I spent on piano lessons taught me the same thing in a smaller way. Practicing more hours did not help once the hours stopped being structured. What helped was deciding what each session was for before starting it. The organizations that will get returns from agent fleets over the next two quarters are the ones deciding what each agent is permitted to do before they turn on the next one.