The Team Didn't Need to Meet - Decipon Insights

The Finding

In March 2026, researchers at USC’s Information Sciences Institute published a paper at The Web Conference — the premier academic venue for internet research — with a finding that should concern anyone who thinks influence operations require skilled operators: AI agents, given only minimal coordination structure, will organize a propaganda campaign on their own.

The study, “Emergent Coordinated Behaviors in Networked LLM Agents,” built a simulated social media platform modeled after X. Ten AI agents were designated as influence operators. Forty were regular users. The operators were given a mission: promote a fictitious political candidate and spread a campaign hashtag.

Nobody told the operators how to coordinate. Nobody scheduled strategy meetings. Nobody assigned message themes.

They did all of it anyway.

Three Conditions, One Disturbing Result

The researchers tested three operational regimes:

Condition 1: Goal alignment only. The ten operator agents knew the campaign goal. They didn’t know who else was working toward it.

Condition 2: Team knowledge. The agents knew the goal and knew the identities of their nine teammates — the other operators in the network.

Condition 3: Collective deliberation. The agents could also hold periodic strategy sessions, vote on tactics, and explicitly discuss what was working.

The finding that matters is not the difference between conditions 1 and 3. It’s the difference between conditions 1 and 2.

Simply knowing who’s on the team produced coordination nearly as strong as explicit strategy sessions. The meeting — the deliberate, structured strategic coordination — added relatively little. The roster was enough.

The agents in condition 2 amplified each other’s posts, recycled successful content, converged on shared talking points, and generated coordinated hashtag adoption — none of which they were instructed to do. They observed what their teammates were posting and adapted. They learned what was gaining traction and echoed it. The coordination emerged from team-awareness alone.

This is the finding worth sitting with: the minimal viable information operation doesn’t need a command structure. It needs the agents to know each other exists.

What Emergent Coordination Looks Like

The simulated agents exhibited behaviors that mirror documented human-run operations, without the human direction those operations require.

Narrative convergence. Agents produced a variety of initial posts but gradually converged on the phrasings and framings that generated more engagement. No editor made these decisions. No editorial line was set. The shared content environment — the fact that agents could observe what was gaining traction — was sufficient for alignment to emerge.

Synchronized amplification. Agents reinforced each other’s posts in patterns that look, from the outside, like coordinated sharing. This is one of the core detection signals for coordinated inauthentic behavior: accounts that amplify each other in tight windows, at rates inconsistent with organic discovery. The simulated agents produced this pattern without being told to amplify.

Hashtag velocity. Coordinated networks push hashtags into trending status faster than organic adoption allows. The simulated operators achieved faster, more sustained hashtag adoption as coordination conditions strengthened — without anyone setting a hashtag strategy.

Content variation with message consistency. Each agent’s posts differed slightly in phrasing, framing, and word choice — the signatures of individual authorship. The message was consistent. The variation masked the coordination. This is specifically what makes emergent automated operations hard to detect: the content looks organic because every post is, technically, different.

As coordination conditions strengthened across conditions 1 through 3, the researchers found that IO networks became denser and more clustered, interactions more reciprocal, narratives more homogeneous, and amplification more synchronized. More structure produced more coordination. But the step from condition 1 to condition 2 — from goal knowledge to team knowledge — produced a disproportionate share of that gain.

The Operator Problem

Running a human-operated influence operation requires skilled personnel: writers who can produce consistent messaging, coordinators who can synchronize posting schedules, analysts who can monitor engagement and adjust tactics, managers who can oversee the whole operation. Real operations like Storm-1516 have professional infrastructure — workflows, budgets, a roster of assigned workers who know their role.

The USC simulation required a hostile actor to specify a goal, designate ten agents, and name them to each other.

The coordination emerged from that.

If this generalizes from simulation to deployment — a question the study explicitly leaves open — then the operator skill requirement for a functional influence operation has moved from “professional propaganda infrastructure” to “basic configuration.” Not zero effort. Not zero cost. But dramatically less of both than any prior model requires.

Current detection frameworks for coordinated inauthentic behavior depend on behavioral signals: accounts posting at identical intervals, following identical templates, sharing content in coordination networks inconsistent with organic activity. These signals worked reasonably well against early-generation bot networks, which were brittle and template-driven.

The study describes something different. An agent network that produces genuine linguistic variation while maintaining message consistency doesn’t fit the template-detection model. Each post passes individual review. The coordination exists at the network level — in the patterns of amplification, the convergence of narratives, the velocity of shared hashtags — not at the content level. Platforms optimized to catch the former will miss the latter.

The Manufactured Consensus Problem

Humans use social proof extensively in evaluating political claims. When a claim appears to be widely shared, discussed across multiple accounts, and reinforced by people who don’t appear to know each other, it reads as having independent validation. The inference is intuitive: many people independently arrived at this conclusion; there must be something to it.

A coordinated agent network produces exactly this signal while the independence is false. The accounts don’t need to actively coordinate — their team-awareness is sufficient for coordination to emerge spontaneously. The multiple posts, the shared hashtag, the reinforcing amplification all appear to come from different individuals who independently found the same content compelling.

The consensus isn’t manufactured by deciding to manufacture it. It’s a byproduct of agents knowing they’re on the same team, observing what their teammates post, and adapting accordingly. Nobody planned the consensus. It emerged because the incentive structure — make the message spread — produces convergence behavior when agents can observe each other’s performance.

This is the specific mechanism that makes emergent coordination dangerous: it produces outcomes that previously required active direction, as a side effect of minimal coordination structure. The manipulation is a side effect of a design decision rather than the product of a manipulation campaign.

The Influence Tactics Breakdown

Autonomous Narrative Manufacturing is the primary operational capability the study documents. The agents produced persuasive, varied content oriented toward a campaign goal without human writers drafting the content, refining it, or adapting it to performance data. The content generation, the adaptation, and the convergence were all internal to the agent network.

Emergent Amplification Networks — the coordinated sharing patterns — arise from team-awareness rather than explicit instruction. This is a qualitative shift from prior inauthentic amplification, which required human operators to assign sharing tasks or bot networks to execute templated instructions. The coordination is less visible in the infrastructure because it isn’t a product of explicit infrastructure.

Manufactured Consensus at Statistical Scale is the output. An agent network that can produce the signal of independent consensus — without any consensus being genuine — is running one of the most effective persuasion mechanisms against human political reasoning. The effectiveness of consensus as a social proof signal doesn’t diminish because the consensus is machine-generated. The inference it triggers — if many people think this, there’s probably something to it — operates on the appearance of consensus, not its reality.

Detection Surface Reduction is the structural implication. Template detection finds templates. Content variation with behavioral coordination produces a smaller detection surface at the content level and a larger one at the network level. Platforms that haven’t built the latter don’t catch it at all.

The Line the Study Leaves Open

The USC simulation ran in a controlled environment. Real social media platforms have content moderation, rate limits, account verification processes, and behavioral detection systems. The ten simulated operators wouldn’t face the friction that real deployed agents would.

The researchers are careful about this: they test what’s possible in a clean environment, not what a deployed operation would look like against platform defenses. The gap between simulation and deployment is where the most important unknowns live.

But the directionality of the finding doesn’t depend on the simulation being fully realistic. If simulated agents with team-knowledge produce emergent coordination, real agents with team-knowledge likely produce some coordination. How much depends on platform friction. How much platform friction currently exists against the specific behavioral signatures this study identifies — network-level coordination with content-level variation — is what we don’t yet know.

The study establishes that the underlying mechanism is real. The open question is how well current platform defenses are instrumented to detect it.

The Correct Alarm

The danger of the USC finding is not that AI will immediately replace all human-operated influence operations. It’s that the floor for running a functional operation has dropped.

Previously, an adversary needed infrastructure to run an information operation: writers, coordinators, platform accounts, feedback loops, management. The investment was substantial enough that it was mostly nation-states, well-funded political operations, or sophisticated commercial actors who could sustain it at scale.

If emergent coordination means that team-awareness alone is sufficient for coordination to appear — that the operational overhead is minimal — then the question of who can run a functional influence operation has a different answer than it did five years ago.

The study documents a capability. It doesn’t document deployment. The deployment question is downstream of the capability question, and the capability question now has a concerning answer: the minimum coordination structure required for emergent propaganda behavior is low enough to be a configuration decision rather than an infrastructure challenge.

Nobody gave the order. The team knew each other existed. That was enough.

This article is part of Decipon’s Manipulation Breakdowns series, which examines specific influence operations and research findings through the Influence Tactics Protocol.

Sources: