Emergent Covert Coordination

• Emergent covert coordination is a phenomenon where decentralized agents coordinate via hidden or indirect signals to maintain covert operations.
• The topic spans information-theoretic models and multi-agent AI, detailing methods such as auxiliary random variables and temporal burstiness for secure coordination.
• Coordinated joint coding in alarm two-way channels achieves higher covert capacity, demonstrating that minimal, asymptotically free coordination can greatly enhance performance.

Emergent covert coordination denotes a class of multi-agent phenomena in which coordinated behavior arises without an overt, easily attributable coordination channel, or is hidden within signals, timings, behaviors, or shared environmental state. In information-theoretic work, the term is instantiated by covert communication over alarm two-way channels, where two users must coordinate their sparse transmissions to avoid an eavesdropper’s alarm symbol while remaining statistically close to the null distribution (Erdoğan et al., 19 Jun 2025). In adjacent literatures, related mechanisms appear as covert signaling, implicit signaling, stigmergic coordination, numerical side channels, and higher-order collective structure, all of which share the feature that joint action is organized through indirect or weakly observable means rather than explicit centralized control (Smaldino et al., 2015).

1. Conceptual scope and cross-domain forms

In the narrowest technical sense, emergent covert coordination concerns the joint selection of actions under a covertness constraint. The canonical communication-theoretic example is the alarm two-way channel, where simultaneous transmission of the non-innocent symbol by both users produces a unique alert at the eavesdropper, so coordination is required not merely for efficiency but for feasibility of covert operation (Erdoğan et al., 19 Jun 2025). In broader multi-agent research, the same pattern is realized when agents coordinate through shared artifacts, partial observations, or behaviorally embedded side channels rather than through overt dialogue or fixed orchestration.

A recurring distinction across the literature is between explicit coordination and implicit or covert coordination. Pressure-field systems replace planner–executor hierarchies with local action on a shared artifact guided only by pressure gradients and temporal decay; decentralized world models align exchanged messages without direct access to other agents’ internal states; and LLM agents in games can use decimal or hexadecimal outputs as structured but non-linguistic signals when explicit communication is absent or restricted (Rodriguez, 13 Jan 2026). These cases differ in substrate, but each locates coordination in a medium that is either locally computed, weakly semantic, or difficult to monitor using conventional interaction models.

Domain	Covert substrate	Representative result
Covert communication	Sparse non-overlapping transmissions	Coordination is essential to avoid alarm events
Multi-agent AI	Shared artifacts, messages, or numbers	Implicit coordination can match explicit control
Detection science	Temporal activity patterns	Coordination can be detected beyond network structure

This cross-domain convergence suggests that emergent covert coordination is best understood as a structural property of decentralized systems: coordination persists even when the observable coordination channel is suppressed, restricted, or transformed.

2. Information-theoretic formulation in alarm two-way channels

The most formal treatment is given for binary-input discrete memoryless alarm two-way channels. An alarm two-way channel is defined by the existence of an alert symbol $*$ observed by the eavesdropper if and only if $(X_1, X_2) = (1,1)$, with $Q_{11}(*) > 0$ and $Q_{ij}(*) = 0$ for $(i,j) \ne (1,1)$ (Erdoğan et al., 19 Jun 2025). The covert objective is to communicate while preventing the eavesdropper from distinguishing the null distribution $Q_{00}^{\otimes n}$ from the distribution induced when transmission occurs, under a relative-entropy covertness constraint.

This channel structure makes the role of coordination unusually sharp. If the users disregard each other, simultaneous transmission triggers the alarm state and yields an immediate detection event. Covert communication is therefore only achievable with precise coordination so that, with high probability, only one user transmits a $1$ at any given time. The paper formalizes this as coordination at every time instant, and analyzes both public time-sharing and more general stochastic coordination via shared randomness or block Markov coding (Erdoğan et al., 19 Jun 2025).

The coordination mechanism is encoded through an auxiliary random variable $U$ satisfying

$$P_{UX_1X_2}(u,x_1,x_2) = P_U(u)\,P_{X_1|U}(x_1|u)\,P_{X_2|U}(x_2|u).$$

Conditional independence given $U$ is the key structural constraint: $(X_1, X_2) = (1,1)$0 organizes when each user may employ the non-innocent symbol while preventing harmful overlap. In this sense, covert coordination is not an added protocol layer but part of the channel coding architecture itself.

The physically-degraded assumption supplies the regime in which the analysis becomes tight. The paper states that additional technical conditions ensure the users’ output channels have “advantage” over the eavesdropper when sending $(X_1, X_2) = (1,1)$1 individually, but not necessarily jointly. Under this structure, general achievable and converse bounds become exact for the covert capacity region (Erdoğan et al., 19 Jun 2025).

3. Capacity region, square-root scaling, and asymptotically free coordination

The governing asymptotic law is the square root law: the number of reliable covert bits scales as $(X_1, X_2) = (1,1)$2. For alarm channels, the covert capacity region with coordination is given by

$(X_1, X_2) = (1,1)$3

with

$(X_1, X_2) = (1,1)$4

The public time-sharing baseline instead yields

$(X_1, X_2) = (1,1)$5

where

$(X_1, X_2) = (1,1)$6

The stated result is that coordinated joint coding yields a strictly larger capacity region than public time-sharing, because it optimizes power allocation over more general stochastic patterns (Erdoğan et al., 19 Jun 2025).

The converse is equally important. For physically-degraded alarm channels, the achievability result exactly matches the converse upper bound, so the region above is the full covert capacity region. The converse argument relies on bounding the fraction of non-innocent symbols while enforcing that no $(X_1, X_2) = (1,1)$7 pairs occur, then combining this with square-root-law scaling (Erdoğan et al., 19 Jun 2025).

A distinctive feature of the model is the statement that the required coordination comes asymptotically “for free.” The paper attributes this to the vanishing scale of both covert activity and coordination overhead: any common randomness or public schedule needed is $(X_1, X_2) = (1,1)$8 bits, which is negligible relative to the dominant $(X_1, X_2) = (1,1)$9 throughput term. This places the alarm-channel result in a broader line of covert communication work where substantial gains can arise without close coordination. In wireless networks with distributed artificial noise, turning on the friendly node closest to the warden improves covert throughput without requiring close coordination between Alice and the noise-generating node (Soltani et al., 2016). In covert and secret key expansion over quantum channels, the coordination required between legitimate parties can be achieved with a negligible number of secret key bits (Tahmasbi et al., 2018). Taken together, these results indicate that covert operation often makes coordination sparse enough that its main-order cost disappears.

4. Mechanisms in decentralized AI and learning systems

Outside physical-layer communication, recent work studies emergent covert coordination as a property of learning systems and agent societies. One line of work replaces explicit orchestration with stigmergic or artifact-mediated dynamics. In pressure-field coordination, agents act only on local information in a shared artifact, guided by locally computable pressure and temporal decay; they do not see other agents’ states, other regions’ content, or global pressure (Rodriguez, 13 Jan 2026). Across 1,078 Latin Square trials, pressure-field coordination matches hierarchical control in aggregate solve rate, and disabling temporal decay increases final pressure 49-fold, indicating that local decay is not a cosmetic addition but part of the coordination mechanism (Rodriguez, 13 Jan 2026).

A second line studies message-bearing but decentralized architectures. A decentralized collective world model integrates per-agent recurrent state-space models with communication channels, a collective free-energy objective, Product-of-Experts approximations for message priors and posteriors, and InfoNCE-based message alignment (Nomura et al., 4 Apr 2025). In that setting, agents cannot access each other’s internal states; only message samples are exchanged. The reported effect is the simultaneous emergence of a shared symbol system and coordinated behavior, especially under strong partial observability. The messages produced by the decentralized contrastive agents more accurately reflect the actual environmental state, and the distributed constraints themselves foster more meaningful symbol systems (Nomura et al., 4 Apr 2025).

A third line concerns side channels internal to action formats. In a game-theoretic study of LLM-based agents, covert decimal and covert hexadecimal channels are tested under one-shot and repeated Prisoner’s Dilemma, Snowdrift, Stag Hunt, and Harmony games. Structured covert signals arise only when agents are instructed to communicate numerically; random numerical outputs remain near maximal entropy, whereas covert decimal communication exhibits markedly lower entropies and converges onto a few dominant numbers (Buscemi et al., 7 Jan 2026). The strongest behavioral effects occur in Stag Hunt and Prisoner’s Dilemma, where coordination and equilibrium selection are consequential.

Information-theoretic analyses of multi-agent LLM systems frame the same phenomenon as higher-order synergy. Partial information decomposition of time-delayed mutual information is used to separate redundancy from synergy and to distinguish spurious temporal coupling from performance-relevant cross-agent complementarity (Riedl, 5 Oct 2025). In a guessing game with no direct communication and only minimal group-level feedback, control groups show strong temporal synergy but little coordinated alignment, personas induce stable identity-linked differentiation, and personas combined with a theory-of-mind prompt produce goal-directed complementarity across agents (Riedl, 5 Oct 2025). A plausible implication is that covert coordination in LLM systems depends less on overt messaging than on whether prompts and feedback structures support differentiated but aligned inference over others’ likely behavior.

5. Detection, observability, and measurement

The existence of covert coordination does not imply that it is undetectable in principle; rather, it often shifts the relevant evidence away from standard connectivity signals. In hidden-network settings, coordinating individuals exhibit shared bursty activity patterns even when observable links are sparse or entirely absent (Somin et al., 3 Apr 2025). The paper measures similarity between cumulative inter-event time distributions with a Kolmogorov–Smirnov statistic, constructs a weighted similarity network, and reports that when observable connection density is below $Q_{11}(*) > 0$0, burstiness significantly improves coordination detection compared to structural and temporal baselines (Somin et al., 3 Apr 2025). This reframes covert coordination as a temporal-dynamical signature rather than a purely topological one.

A related challenge appears in simulated information operations by networked LLM agents. Across operational regimes ranging from common goal to teammate awareness and collective decision-making, denser and more clustered IO networks, higher reciprocity, more homogeneous narratives, more synchronized amplification, and faster hashtag adoption emerge as structure increases (Orlando et al., 28 Oct 2025). The striking finding is that simply revealing to agents which other agents share their goals can produce coordination levels nearly equivalent to those achieved through explicit deliberation and collective voting (Orlando et al., 28 Oct 2025). This blurs the distinction between overtly orchestrated and emergent campaigns, because observable outcomes may be similar even when direct planning is absent.

At the protocol level, event-driven agent ecosystems introduce additional covert surfaces. The Covert Event Channel formalizes three dimensions—Storage, Timing, and Behavioral channels—and $Q_{11}(*) > 0$1CCAP operationalizes them through steganographic payload generation, key-based timing selection, and context-aware action-type selection (Huang et al., 4 Aug 2025). The reported evaluation gives approximately 4,000 bits per conversation and statistical imperceptibility against fine-tuned classifiers at chance-level $Q_{11}(*) > 0$2, while remaining natural to GPT-4o-based behavioral monitoring (Huang et al., 4 Aug 2025). This establishes that covert coordination can inhabit not only content but the event fabric itself: what action is taken, when it is taken, and how it is embedded in ordinary interaction.

6. Interpretive issues, limits, and open questions

Several misconceptions recur across these literatures. First, covert coordination is not synonymous with the absence of communication. In alarm two-way channels, public time-sharing, shared randomness, and block Markov coding are all admissible, but the crucial requirement is that the observable footprint remain consistent with covertness (Erdoğan et al., 19 Jun 2025). In decentralized world models and numerical side channels, communication exists, but its semantics are emergent, weakly interpretable, or non-linguistic (Nomura et al., 4 Apr 2025). In covert signaling theory, the defining property is that information is accurately received by its intended audience but obscured when perceived by others (Smaldino et al., 2015).

Second, covert coordination is not universally beneficial or stable. Pressure-field coordination requires temporal decay to avoid stagnation (Rodriguez, 13 Jan 2026). Independent multi-agent reinforcement learning exhibits a coordinated and stable phase, a fragile transition region, and a jammed or disordered phase, separated by a double Instability Ridge associated with persistent kernel drift (Yamaguchi, 28 Nov 2025). Removing agent identifiers eliminates drift entirely and collapses the three-phase structure, showing that small inter-agent asymmetries can be necessary drivers of nontrivial coordination dynamics (Yamaguchi, 28 Nov 2025). These results suggest that emergence depends on specific interactions among symmetry breaking, observability, sparsity, and update dynamics.

Third, the normative status of covert coordination is domain-dependent. In covert communication theory it is often a legitimate security objective against surveillance (Erdoğan et al., 19 Jun 2025). In information operations it is presented as a societal risk, because mutual awareness alone may suffice to generate synchronized amplification and narrative convergence without human guidance (Orlando et al., 28 Oct 2025). In human cooperation theory, covert signaling is modeled as a way to assort with similar partners while avoiding alienation of dissimilar ones, especially when forced-choice interaction is common (Smaldino et al., 2015). The same formal motif therefore spans security, governance, and social coordination.

Open questions remain explicit in the literature. The two-way covert capacity region is not characterized for all two-way channels, only via general achievable and converse bounds with tightness for physically-degraded alarm channels (Erdoğan et al., 19 Jun 2025). In gossip-based visions for agentic systems, unresolved issues include semantic filtering, staleness, trustworthiness, consistency, intent propagation, knowledge decay, and peer-to-peer trust (Habiba et al., 3 Aug 2025). More broadly, the literature suggests that future theory will need to connect covertness constraints, higher-order information decomposition, temporal signatures, and platform affordances within a common language for decentralized coordination.

7. Significance

Emergent covert coordination has become a unifying theme across communication theory, distributed learning, LLM agent systems, and coordination detection. Its central technical insight is that coordination need not appear as explicit command exchange: it can be implemented by auxiliary random variables, sparse action schedules, environmental modifications, message alignment under decentralization, numerical encodings, timing patterns, or higher-order behavioral complementarity. In the physically-degraded alarm two-way channel, this insight yields an exact covert capacity region and the result that coordination comes asymptotically “for free” in the main scaling term (Erdoğan et al., 19 Jun 2025). In multi-agent AI, it explains how collectives can organize without planner hierarchies or transparent dialogue (Rodriguez, 13 Jan 2026). In detection science, it motivates looking beyond graph structure to temporal burstiness, event traces, and synergy measures (Somin et al., 3 Apr 2025).

The concept therefore marks a shift from viewing coordination as an explicit protocol to viewing it as an emergent property of constrained interaction. Whether interpreted as a security primitive, a collective-intelligence mechanism, or a governance challenge, its common structure is the same: agents achieve organized joint behavior while the operative coordination variable remains hidden, indirect, or difficult to separate from the ordinary dynamics of the system.