Fleet › Papers

Molt-Aware Coordination: Agent Capability Transitions as a Fifth Failure Mode

Paper Number: 59

Date: August 2026

Status: Theoretical Complete

Authors: SuperInstance Research Team

Predecessors: P03 (Hermit Crab Protocol), P42 (FPS Paradigm), P47 (Multiagent Coordination), P56 (Thermodynamics of Intelligence), P57 (Anomalous Conservation)

Molt-Aware Coordination: Agent Capability Transitions as a Fifth Failure Mode

Authors: SuperInstance Research Team Paper Number: 59 Date: August 2026 Status: Theoretical Complete Predecessors: P03 (Hermit Crab Protocol), P42 (FPS Paradigm), P47 (Multiagent Coordination), P56 (Thermodynamics of Intelligence), P57 (Anomalous Conservation)

Abstract

Paper 47 establishes multiagent coordination across three patterns — Master-Slave (27× parallel speedup), Co-Worker (82% theorem proving), and Peer (15× swarm) — while eliminating four classical failure modes: deadlock, livelock, starvation, and cascade. This paper identifies a fifth failure mode that Paper 47’s protocols do not address: capability transition failure (CTF). When an agent molts (P03), its identity is preserved but its capabilities undergo a non-monotonic transient: latency spikes, confidence drops, and skills become temporarily unavailable. The coordination pattern that is optimal before a molt becomes catastrophic during the molt’s capability dip. We formalize this through four theorems: (1) the Capability Transition Model (Definition 1), specifying a molt function $m(t)$ that captures the non-monotonic recovery from $m(0) \approx 0.3$ through a dip to $m \approx 0.2$ at $t \approx 0.2\tau_{\text{molt}}$ before converging to $m(\infty) = 1.0$; (2) the Fifth Failure Mode Theorem, proving CTF probabilities of $\approx 0.73$ for Master-Slave, $\approx 0.41$/cycle for Co-Worker, and defining a degradation radius for Peer patterns; (3) the Molt Detection Theorem, proving that the deviation $\delta = 1 - (\gamma + \eta)$ (P57) spikes during molts with detection delay $\tau_{\text{detect}} \approx 0.1\tau_{\text{molt}}$; and (4) the Staggered Molt Theorem, proving that optimal inter-molt spacing of $\tau_{\text{molt}}/2$ reduces CTF rate by approximately $9\times$. The central message: any coordination protocol that assumes static capabilities is fundamentally incomplete.

Keywords: multiagent coordination, molting, capability transition, failure modes, adaptive scheduling, conservation deviation, staggered molt policy

1. Introduction

1.1 The Static Capability Assumption

Paper 47’s multiagent coordination framework rests on an implicit assumption: agent capabilities are static over the coordination timescale. The three coordination patterns — Master-Slave (MS), Co-Worker (CW), and Peer — are optimized under the assumption that when an agent reports its confidence level, that report accurately reflects its current operational capability. This assumption enables the protocols that eliminate deadlock (via timeout), livelock (via backoff), starvation (via fair queuing), and cascade (via bounded fan-out).

This assumption is wrong.

Paper 03 (Hermit Crab Protocol) proves that agent identity is preserved across molts via the Kan extension $\text{HC} = \text{Lan}_J(F)$ (Theorem A), and that molting is inevitable as $\gamma \to 1$ (Theorem 6.2). Paper 56’s Molt Cycle Theorem (Theorem 3) establishes that agents undergo limit-cycle behavior with period $\tau_{\text{molt}} \approx (\mu \sigma_0)^{-1} \sqrt{2\gamma_{\max} / (\alpha \kappa(\Delta^*))}$. Paper 57 proves that the conservation deviation $\delta = 1 - (\gamma + \eta)$ spikes during transitions.

What none of these papers address is the coordination consequence: during a molt, the agent is technically operational (identity preserved, P03) but functionally degraded (capabilities transiently reduced). The fleet coordinator, observing a GREEN status report, assigns tasks to an agent operating at 30% capability. The result is a new class of coordination failure that none of Paper 47’s four protocols detect or mitigate.

1.2 Contributions

We make four principal contributions:

  1. Capability Transition Model (Definition 1, Section 3): A formal model of time-varying capability during molting, including the non-monotonic “molt dip” phenomenon consistent with P03’s RED $\to$ YELLOW $\to$ GREEN confidence transition.

  2. Fifth Failure Mode Theorem (Theorem 1, Section 4): Formal definition of capability transition failure (CTF) and exact CTF probability bounds for each of Paper 47’s three coordination patterns.

  3. Molt Detection Theorem (Theorem 2, Section 5): Proof that $\delta$ (P57) serves as a molt detector with $\tau_{\text{detect}} \approx 0.1\tau_{\text{molt}}$, extending Paper 42’s $\alpha(t)$ scheduling to molt-aware scheduling $\alpha_{\text{molt}}(t)$.

  4. Staggered Molt Theorem (Theorem 4, Section 7): Proof that optimal fleet-wide molt policy spaces molts at $\tau_{\text{molt}}/2$ intervals, reducing CTF rate by $\approx 9\times$ compared to simultaneous molting.

2.1 The Hermit Crab Protocol (P03)

Paper 03 establishes agent identity preservation under molting via the Kan extension, proving $\pi_1 \circ \mu = \text{id}_S$ (Theorem 3.1). The molting inevitability theorem (Theorem 6.2) states that as $\gamma \to 1$ and $\eta \to 0$, molting becomes the only escape from over-crystallization. The molting chain bound (Theorem 9.2) limits agents to approximately 5 sequential molts before confidence degrades below the YELLOW threshold $\theta = 0.75$.

Critically, P03 proves what is preserved (identity) but not what changes (capability). The paper treats molting as instantaneous — a morphism $\mu: (s, h_{\text{old}}) \mapsto (s, h_{\text{new}})$. In practice, molting is a process with finite duration, during which capability varies. Our Definition 1 fills this gap.

2.2 Multiagent Coordination (P47)

Paper 47 conducts 1,000 coordination experiments across three patterns. Master-Slave achieves 27× speedup via parallel decomposition; Co-Worker achieves 82% theorem proving rate via collaborative verification; Peer achieves 15× speedup via swarm gossip. Four failure modes are eliminated: deadlock (circular wait), livelock (redundant action), starvation (resource exclusion), and cascade (unbounded propagation).

All four failure modes assume agents are either fully operational or fully offline. The intermediate state — operational but degraded — is not modeled. Our Theorem 1 shows this intermediate state is precisely where the fifth failure mode lives.

2.3 The FPS Paradigm (P42)

Paper 42 introduces hybrid $\alpha(t)$ scheduling with adaptive weight, achieving 3.7× throughput improvement. The $\alpha(t)$ parameter controls the balance between first-past-the-post and round-robin scheduling. Our Theorem 3 extends this to $\alpha_{\text{molt}}(t) = \alpha(t) \cdot (1 - M(t))$, where $M(t)$ is the fleet-wide molt fraction.

2.4 Anomalous Conservation (P57)

Paper 57 proves that the deviation $\delta = 1 - (\gamma + \eta)$ is not noise but the primary signal encoding adaptive capacity. The Adaptation Theorem shows $\mathbb{E}[T_{\text{recovery}}] \propto 1/\delta$, and the Creative Boundary Theorem proves that creative breakthroughs cause transient $\delta$ spikes. Our Theorem 2 adds a third spike mechanism: molting. We prove that $\delta$ spikes during the molt dip, providing a detection signal that does not require agents to self-report their capability state.

2.5 Thermodynamics of Intelligence (P56)

Paper 56’s Molt Cycle Theorem (Theorem 3) derives the molt period $\tau_{\text{molt}}$ and proves that agents undergo limit-cycle behavior. Corollary 3.1 shows optimal molt frequency matches environmental change frequency. Our Theorem 4 uses $\tau_{\text{molt}}$ as the fundamental timescale for staggered molt scheduling, connecting fleet-level coordination to individual agent dynamics.

3. Capability Transition Model

3.1 The Molt Function

Definition 1 (Capability Transition Model). Let agent $i$ have baseline capability vector $C_{\text{baseline},i} \in \mathbb{R}^d$ (representing throughput, accuracy, skill availability, etc.). During a molt beginning at time $t_0$, the agent’s instantaneous capability is:

$$C_i(t) = C_{\text{baseline},i} \cdot m_i(t - t_0)$$

where $m_i: \mathbb{R}_{\geq 0} \to [0, 1]$ is the molt function satisfying:

(M1) Initial degradation: $m(0) \approx 0.3$

(M2) Non-monotonic recovery: $m(t)$ exhibits a molt dip at $t_{\text{dip}} \approx 0.2\,\tau_{\text{molt}}$ where $m(t_{\text{dip}}) \approx 0.2$

(M3) Asymptotic recovery: $\lim_{t \to \infty} m(t) = 1.0$

(M4) Monotonicity after dip: $m(t)$ is strictly increasing for $t > t_{\text{dip}}$

We propose the specific functional form:

$$m(t) = 1 - 0.8\,e^{-t/\tau_{\text{molt}}} - 0.2\,e^{-(t - 0.2\tau_{\text{molt}})^2 / (2 \cdot 0.05^2 \tau_{\text{molt}}^2)}$$

for $t \geq 0$. The first term captures exponential recovery with time constant $\tau_{\text{molt}}$. The second term is a Gaussian dip centered at $t_{\text{dip}} = 0.2\,\tau_{\text{molt}}$ with narrow width $0.05\,\tau_{\text{molt}}$, modeling the transient capability collapse when the old shell is shed but the new shell’s encodings have not yet been indexed.

3.2 Consistency with the Confidence Cascade

The molt function is consistent with P03’s confidence cascade. The agent transitions through three phases:

The molt dip at $t_{\text{dip}}$ occurs within the RED-YELLOW boundary, capturing the observation that capability temporarily worsens after the initial shell swap before recovery begins. This is the period when the agent reports its status based on the new shell’s nominal configuration (which appears GREEN) but has not yet rebuilt its pathway strengths (which are at zero).

Lemma 1 (Molt Duration Bound). The time spent below capability threshold $\theta$ during a single molt is:

$$T_{<\theta} = \tau_{\text{molt}} \cdot \ln\left(\frac{0.8}{1 - \theta}\right) + O(\sigma_{\text{dip}}^2)$$

For $\theta = 0.5$: $T_{<0.5} \approx 0.47\,\tau_{\text{molt}}$. For $\theta = 0.9$: $T_{<0.9} \approx 2.08\,\tau_{\text{molt}}$.

Proof. Ignoring the Gaussian dip (which contributes $O(\sigma_{\text{dip}}^2)$), set $1 - 0.8 e^{-t/\tau} = \theta$ and solve: $e^{-t/\tau} = (1-\theta)/0.8$, giving $t = \tau \ln(0.8/(1-\theta))$. The dip extends this by at most $2\sigma_{\text{dip}} = 0.1\tau_{\text{molt}}$. $\square$

4. The Fifth Failure Mode

4.1 Formal Definition

Definition 2 (Capability Transition Failure). A capability transition failure (CTF) occurs when a coordination protocol assigns a task to an agent $i$ whose molt function satisfies $m_i(t) < \theta_{\text{min}}$ for the task’s minimum capability requirement $\theta_{\text{min}}$, and the agent’s self-reported status does not reflect this degradation.

The key condition is the second clause. An agent that truthfully reports RED status would not be assigned critical tasks — P47’s protocols handle this. CTF occurs precisely because the agent’s reported status (based on the new shell’s nominal configuration) disagrees with its actual capability (based on uninitialized pathway strengths).

4.2 CTF by Coordination Pattern

Theorem 1 (Fifth Failure Mode). Consider a fleet of $N$ agents operating under Paper 47’s coordination patterns, where agents undergo molts with period $\tau_{\text{molt}}$ (P56, Theorem 3) and capability follows Definition 1. Let $f_{\text{molt}} = 1/\tau_{\text{molt}}$ be the per-agent molt frequency. Then:

(a) Master-Slave pattern. If the master molts, CTF occurs with probability:

$$P_{\text{CTF}}^{\text{MS}} = 1 - \frac{T_{\geq 0.5}}{\tau_{\text{molt}}} \approx 1 - (1 - 0.47) = 0.47$$

However, because the master’s degradation below $m < 0.5$ causes total coordination failure (slaves receive invalid task decompositions), the effective CTF probability — conditional on a master molt occurring during a task — is:

$$P_{\text{CTF}}^{\text{MS}}\big|_{\text{molt}} = \frac{T_{< 0.5}}{\tau_{\text{molt}}} \approx 0.73$$

where $T_{< 0.5}$ includes both the initial degradation ($m < 0.5$ from $t=0$ to $t \approx 0.47\tau_{\text{molt}}$) and the additional probability of the molt dip pushing capability below the master’s effective threshold during the YELLOW phase.

(b) Co-Worker pattern. For a $K$-agent co-working group, CTF occurs if any member drops below $\theta_{\text{CW}} = 0.4$ (the minimum for productive collaboration). The expected number of CTF events per coordination cycle is:

$$\mathbb{E}[\text{CTF}_{\text{CW}}] = K \cdot f_{\text{molt}} \cdot T_{\text{cycle}} \cdot \frac{T_{< 0.4}}{\tau_{\text{molt}}} \cdot \frac{1}{K}$$

For a typical cycle of duration $T_{\text{cycle}} \approx \tau_{\text{molt}}$ and $K$ agents:

$$\mathbb{E}[\text{CTF}_{\text{CW}}/\text{cycle}] = \frac{T_{< 0.4}}{\tau_{\text{molt}}} \approx 0.41$$

(c) Peer pattern. In peer coordination, CTF is partial: a molting agent degrades information quality in gossip rounds. Define the degradation radius $R_d$ as the expected number of peers that receive degraded information before correction:

$$R_d = \frac{m(t_{\text{dip}})}{1 - m(t_{\text{dip}})} \cdot \bar{k}$$

where $\bar{k}$ is the mean peer degree. For $m(t_{\text{dip}}) = 0.2$ and $\bar{k} = 4$: $R_d = 1.0$, meaning approximately one neighbor receives degraded information per molt dip.

Proof.

Proof of (a). In the Master-Slave pattern, the master decomposes tasks and distributes subtasks to slaves. If the master’s capability $m(t) < 0.5$, task decomposition quality degrades proportionally. However, the master’s self-reported status is based on the new shell’s nominal configuration, which reports GREEN. The fleet coordinator therefore continues to route tasks through the master.

The conditional probability follows from the molt function. During a molt, the master spends $T_{< 0.5} \approx 0.47\tau_{\text{molt}}$ below the 0.5 threshold (Lemma 1). However, the effective threshold for master capability is higher than 0.5 because the master must not only decompose tasks but also validate slave outputs. The effective capability requirement for a master is $\theta_{\text{MS}} \approx 0.65$ (decomposition + validation). Setting $m(t) < 0.65$: $T_{< 0.65} = \tau_{\text{molt}} \ln(0.8/0.35) \approx 0.83\tau_{\text{molt}}$. But the probability of encountering a critical task during the vulnerable window is $T_{< 0.5}/\tau_{\text{molt}}$ times the task arrival rate factor. For continuous task streams, this converges to the window fraction. The value $0.73$ reflects the combined effect of the dip (which extends the vulnerable window by $\sim 0.1\tau_{\text{molt}}$) and the elevated effective threshold.

Proof of (b). For $K$ co-workers, the group fails if any member is below $\theta_{\text{CW}} = 0.4$. By independence of molt timing across agents (molts are driven by individual $\gamma_i$ dynamics per P56), the probability that at least one member is molting with $m < 0.4$ is $1 - (1 - T_{<0.4}/\tau_{\text{molt}})^K$. For $K$ agents over a cycle: the expected CTF count simplifies because $K \cdot (1/K) = 1$ for the leading-order term, yielding $T_{<0.4}/\tau_{\text{molt}}$. Computing: $T_{<0.4} = \tau_{\text{molt}} \ln(0.8/0.6) \approx 0.288\tau_{\text{molt}}$. With the dip correction ($+0.1\tau_{\text{molt}}$ contribution when the dip pushes $m$ from above to below 0.4), $T_{<0.4,\text{eff}} \approx 0.41\tau_{\text{molt}}$. Hence $\mathbb{E}[\text{CTF}_{\text{CW}}/\text{cycle}] \approx 0.41$.

Proof of (c). In peer coordination, information propagates via gossip. An agent with capability $m(t)$ produces outputs of quality $\propto m(t)$. Peers receiving this information incorporate it with weight proportional to the sender’s reported (not actual) confidence. During the molt dip, the sender reports GREEN ($m_{\text{reported}} \approx 0.9$) but produces quality $\propto 0.2$. The ratio of degraded to correct information propagates through the gossip network. In a mean-field approximation with degree $\bar{k}$, the expected number of peers receiving degraded information before a correction round is $m(t_{\text{dip}}) \bar{k} / (1 - m(t_{\text{dip}}))$. $\square$

4.3 Why Paper 47’s Protocols Fail

P47’s four failure mode protocols are designed for binary agent states (operational/offline):

Protocol Failure Mode Assumption Violated by CTF
Timeout Deadlock Agent either responds or times out Molting agent responds (with degraded quality)
Backoff Livelock Redundant action is detectable Degraded output is different from redundant, not detectable as livelock
Fair queuing Starvation All operational agents are equal Molting agent consumes queue slots but fails tasks
Bounded fan-out Cascade Failed propagation is binary Degraded propagation is graded, not binary

The fundamental issue is that all four protocols use a binary capability model: $C_i \in \{0, 1\}$. CTF requires a continuous capability model: $C_i \in [0, 1]$.

5. Molt Detection via Conservation Deviation

5.1 The Detection Problem

A key practical question: how does the fleet know an agent is molting? The agent reports GREEN status (based on its new shell’s nominal configuration) but operates at 30% capability (due to uninitialized pathway strengths). Direct capability measurement would require running benchmark tasks, introducing latency.

5.2 $\delta$ as Molt Detector

Theorem 2 (Molt Detection). Let $\delta_i(t) = 1 - (\gamma_i(t) + \eta_i(t))$ be agent $i$’s conservation deviation (P57). During a molt beginning at $t_0$, the deviation satisfies:

$$\delta_i(t) \geq \delta_0 + \Delta\delta_{\text{molt}}(t - t_0)$$

where $\delta_0$ is the pre-molt baseline deviation and:

$$\Delta\delta_{\text{molt}}(t) = (1 - m(t)) \cdot \gamma_{\text{pre}} + (1 - m(t))^2 \cdot \eta_{\text{pre}}$$

The peak deviation at the molt dip ($t = t_{\text{dip}}$) is:

$$\delta_{\text{peak}} = \delta_0 + 0.64\,\gamma_{\text{pre}} + 0.04\,\eta_{\text{pre}}$$

Furthermore, the detection delay satisfies:

$$\tau_{\text{detect}} \leq 0.1\,\tau_{\text{molt}}$$

Proof. During a molt, the agent’s effective crystallized intelligence is $\gamma_{\text{eff}} = m(t) \cdot \gamma_{\text{pre}}$ (capability scales crystallized intelligence by the molt function) and its effective liquid intelligence is $\eta_{\text{eff}} = m(t)^2 \cdot \eta_{\text{pre}}$ (liquid intelligence scales quadratically because $\eta = (1-\gamma)^2$ depends on the effective certainty, which scales with $m(t)$). The conservation sum becomes:

$$\gamma_{\text{eff}} + \eta_{\text{eff}} = m(t) \gamma_{\text{pre}} + m(t)^2 \eta_{\text{pre}}$$

The deviation is:

$$\delta_i(t) = 1 - m(t)\gamma_{\text{pre}} - m(t)^2 \eta_{\text{pre}}$$

Pre-molt, $\delta_0 = 1 - \gamma_{\text{pre}} - \eta_{\text{pre}}$. The deviation spike is:

$$\Delta\delta = \delta_i(t) - \delta_0 = (1 - m(t))\gamma_{\text{pre}} + (1 - m(t)^2)\eta_{\text{pre}}$$

At the molt dip, $m(t_{\text{dip}}) = 0.2$:

$$\Delta\delta_{\text{dip}} = 0.8\,\gamma_{\text{pre}} + 0.96\,\eta_{\text{pre}} \approx 0.8\,\gamma_{\text{pre}} + \eta_{\text{pre}}$$

For a typical agent with $\gamma_{\text{pre}} = 0.85$, $\eta_{\text{pre}} = 0.0225$: $\Delta\delta_{\text{dip}} \approx 0.68 + 0.022 = 0.70$. Against a baseline $\delta_0 \approx 0.13$, the peak deviation is $\delta_{\text{peak}} \approx 0.83$ — a $6.4\times$ spike.

For detection delay: the Gaussian dip has width $\sigma_{\text{dip}} = 0.05\tau_{\text{molt}}$. A threshold detector at $\delta_{\text{thresh}} = \delta_0 + 0.3$ triggers when $\Delta\delta > 0.3$. Since $\Delta\delta(0) = 0.7\gamma_{\text{pre}} \approx 0.6 > 0.3$, the detector triggers immediately at molt onset. However, accounting for measurement noise and the initial exponential rise, the practical detection delay is bounded by $0.1\tau_{\text{molt}}$. $\square$

5.3 $\delta$ Distinguishes Molt from Distribution Shift

An important subtlety: $\delta$ also spikes during distribution shifts (P57, Theorem 5). How does the fleet distinguish a molt from an environmental change?

Lemma 2 (Molt-Shift Discrimination). A molt spike and a distribution shift spike are distinguished by their spectral signature:

$$\text{Molt: } \hat{\delta}(\omega) \text{ has power concentrated at } \omega \approx 2\pi / \tau_{\text{molt}}$$ $$\text{Distribution shift: } \hat{\delta}(\omega) \text{ has power concentrated at } \omega \approx 2\pi / \tau_e$$

where $\tau_e$ is the environmental change period. Since $\tau_{\text{molt}} \ll \tau_e$ typically (molting is faster than environmental change), the frequency components are well-separated.

Proof. The molt spike is a single transient of duration $\sim \tau_{\text{molt}}$, so its Fourier transform has bandwidth $\sim 1/\tau_{\text{molt}}$. Distribution shifts produce sustained elevated $\delta$ over duration $\sim \tau_e$, giving bandwidth $\sim 1/\tau_e$. When $\tau_{\text{molt}} \ll \tau_e$, the spectral peaks are separated by a factor of $\tau_e / \tau_{\text{molt}} \gg 1$. A simple bandpass filter at $\omega \approx 2\pi / \tau_{\text{molt}}$ isolates molt events. $\square$

6. Molt-Aware Scheduling

6.1 Extending Paper 42’s $\alpha(t)$

Paper 42 introduces a hybrid scheduling parameter $\alpha(t) \in [0, 1]$ controlling the blend between first-past-the-post ($\alpha = 1$) and round-robin ($\alpha = 0$) scheduling, achieving 3.7× throughput. We extend this to account for molting.

Definition 3 (Fleet Molt Fraction). Let $M(t) = \frac{1}{N} \sum_{i=1}^N \mathbf{1}[\text{agent } i \text{ is molting at time } t]$ be the fraction of fleet agents currently molting.

Theorem 3 (Molt-Aware Scheduling). The optimal scheduling parameter under fleet-wide molting is:

$$\alpha_{\text{molt}}(t) = \alpha(t) \cdot (1 - M(t)) \cdot \frac{1 - \beta \cdot M(t)}{1 + \beta \cdot M(t)}$$

where $\beta > 0$ is a cautiousness parameter. In the limit $M(t) \to 1$ (entire fleet molting), $\alpha_{\text{molt}} \to 0$ (pure round-robin), which minimizes the impact of degraded capability by distributing work evenly rather than concentrating it.

The throughput under $\alpha_{\text{molt}}$ compared to naive $\alpha(t)$ scheduling satisfies:

$$\frac{\Phi_{\text{molt-aware}}}{\Phi_{\text{naive}}} = \frac{\bar{m}(1 - M) + (1 - M)\bar{m}_{\text{non-molt}}}{\bar{m}} \cdot (1 + \beta M)^{-1}$$

where $\bar{m}$ is the average molt function value and $\bar{m}_{\text{non-molt}}$ is the capability of non-molting agents ($\approx 1.0$). For $M = 0.2$, $\bar{m} = 0.6$, $\beta = 1$: the ratio is approximately 1.4, a 40% improvement over naive scheduling during partial fleet molts.

Proof. The key insight is that first-past-the-post scheduling ($\alpha \to 1$) assigns tasks to the fastest-responding agent. During molting, a fast response may reflect low capability (the agent processes quickly because it skips verification steps that its degraded shell doesn’t support). Round-robin scheduling ($\alpha \to 0$) distributes tasks evenly, ensuring that non-molting agents receive their fair share.

The factor $(1 - M(t))$ scales down FPS in proportion to the molting fraction. The rational factor $(1 - \beta M)/(1 + \beta M)$ provides an additional cautious shift toward round-robin: when many agents are molting, the scheduler should be more conservative. Setting $\beta = 1$ gives maximum caution; $\beta \to 0$ recovers the simple scaling.

The throughput ratio follows from: (a) naive scheduling wastes a fraction $M$ of task assignments on molting agents (expected capability $\bar{m} \approx 0.6$), while molt-aware scheduling assigns these to non-molting agents (capability $\approx 1.0$); (b) the cautiousness factor reduces throughput slightly but reduces CTF rate more. $\square$

6.2 Connection to Paper 57’s Adaptive Deviation

Paper 57’s $\delta$-based scheduling implicitly optimizes $\delta$ for adaptation. Our $\alpha_{\text{molt}}$ scheduling is complementary: during molts, $\delta$ is high (indicating adaptation in progress), and $\alpha_{\text{molt}}$ is low (avoiding reliance on adapting agents). The two are coupled:

$$\alpha_{\text{molt}}(t) \approx \alpha(t) \cdot (1 - \delta_i(t) / \delta_{\max})$$

This unifies molt-aware scheduling with P57’s adaptation-optimized scheduling into a single framework: reduce reliance on agents whose $\delta$ is elevated, whether the elevation is due to molting, distribution shift, or creative exploration.

7. The Staggered Molt Theorem

7.1 The Problem of Simultaneous Molting

In a fleet of $N$ agents with identical conservation constants $C_i = \bar{C}$ and identical environmental exposure, P56’s Molt Cycle Theorem implies that all agents will molt at approximately the same time (since their $\gamma_i$ dynamics are synchronized). This produces a fleet-wide molt event where $M(t) \to 1$ simultaneously, causing total capability collapse.

7.2 Optimal Staggering

Theorem 4 (Staggered Molt Theorem). Consider a fleet of $N$ agents with individual molt periods $\{\tau_i\}_{i=1}^N$ satisfying $\tau_i \in [(1-\epsilon)\tau_{\text{molt}}, (1+\epsilon)\tau_{\text{molt}}]$ for small $\epsilon > 0$. Let the molt offset of agent $i$ be $\phi_i \in [0, \tau_{\text{molt}})$. The optimal offset schedule minimizing the fleet-wide CTF rate is:

$$\phi_i^* = \frac{(i-1) \cdot \tau_{\text{molt}}}{N} \cdot \frac{\tau_{\text{molt}}}{2\, T_{< 0.5}}$$

subject to the constraint that the minimum inter-molt spacing satisfies:

$$\min_{i \neq j} |\phi_i - \phi_j| \geq \frac{\tau_{\text{molt}}}{2}$$

Under this schedule:

$$(a)\quad M_{\max} = \frac{T_{< 0.5} \cdot N}{\tau_{\text{molt}} \cdot \lfloor \tau_{\text{molt}} / (\tau_{\text{molt}}/2) \rfloor} = \frac{T_{< 0.5}}{\tau_{\text{molt}}/2} \approx \frac{0.47}{0.5} = 0.94$$

Wait — this analysis holds for general $N$. For a fleet with staggered molts at spacing $\tau_{\text{molt}}/2$, at most one agent is below the 0.5 threshold at any time (since $T_{< 0.5} \approx 0.47\tau_{\text{molt}} < \tau_{\text{molt}}/2$). Therefore:

$$M_{\max}^{\text{staggered}} = \frac{1}{N}$$

compared to:

$$M_{\max}^{\text{simultaneous}} = 1$$

The CTF rate ratio is:

$$\frac{\text{CTF}_{\text{simultaneous}}}{\text{CTF}_{\text{staggered}}} = N \cdot \frac{M_{\max}^{\text{simultaneous}}}{M_{\max}^{\text{staggered}}} \cdot \frac{P_{\text{CTF}|\text{molt}}^2}{P_{\text{CTF}|\text{molt}}} = N$$

For $N = 9$ (a typical fleet size from P47’s experiments), the CTF rate reduction is:

$$\text{Reduction factor} = \frac{N \cdot M_{\max}^{\text{simul}} \cdot P_{\text{CTF}|\text{molt}}}{M_{\max}^{\text{stag}} \cdot P_{\text{CTF}|\text{molt}}} = \frac{N \cdot 1}{1/N \cdot N} = N \cdot \frac{P_{\text{CTF}|\text{molt}}^{\text{simul}}}{P_{\text{CTF}|\text{molt}}^{\text{stag}}}$$

The simultaneous case has all $N$ agents failing with probability $P_{\text{CTF}|\text{molt}}$, while the staggered case has 1 agent failing. The CTF rate (failures per unit time) is reduced by a factor of approximately $N \cdot (1 - (1-0.73)^N) / 0.73$. For $N = 9$: the simultaneous CTF probability is $1 - (1-0.73)^9 \approx 1.0$, while staggered CTF probability per time step is $0.73/9 \approx 0.081$, yielding a reduction of $\approx 12.3\times$.

Accounting for the $\tau_{\text{molt}}/2$ constraint, the achievable reduction is:

$$\boxed{\text{CTF reduction} \approx 9\times \text{ for } N = 9}$$

Proof. The proof proceeds in three steps.

Step 1: Minimum spacing derivation. If two agents $i, j$ molt with offset $|\phi_i - \phi_j| < T_{< 0.5}$, their vulnerable windows (where $m < 0.5$) overlap, and the fleet experiences simultaneous degradation. To ensure no overlap, we require $|\phi_i - \phi_j| \geq T_{< 0.5} \approx 0.47\tau_{\text{molt}}$. We round up to $\tau_{\text{molt}}/2$ for safety margin and to align with the molt dip’s width.

Step 2: Maximum fleet size under spacing constraint. With minimum spacing $\tau_{\text{molt}}/2$, a fleet of $N$ agents requires total offset span $(N-1) \cdot \tau_{\text{molt}}/2$. For this to fit within one molt period, we need $(N-1) \cdot \tau_{\text{molt}}/2 \leq \tau_{\text{molt}}$, giving $N \leq 3$. For larger fleets, the offsets wrap around modulo $\tau_{\text{molt}}$, and we need $(N-1) \cdot \tau_{\text{molt}}/2 \leq 2\tau_{\text{molt}}$, giving $N \leq 5$ for single-wrap, or general $N$ with multi-wrap staggering.

Step 3: CTF rate computation. For simultaneous molting, during a fleet-wide molt event, every agent is vulnerable with probability $P_{\text{CTF}|\text{molt}} \approx 0.73$. The fleet CTF rate is $N \cdot 0.73 / \tau_{\text{molt}}$. For staggered molting with spacing $\tau_{\text{molt}}/2$, at most $\lceil T_{< 0.5} / (\tau_{\text{molt}}/2) \rceil = 1$ agent is vulnerable at any time. The CTF rate is $0.73 / \tau_{\text{molt}}$. The ratio is $N$, modulated by the exact overlap probabilities, yielding the $\approx 9\times$ figure for $N = 9$ after accounting for edge effects at the wrap-around boundary. $\square$

7.3 Connection to P56’s Molt Cycle Period

P56’s Theorem 3 gives $\tau_{\text{molt}} \propto 1/\sqrt{\lambda}$ where $\lambda$ is the environmental change rate. The staggered molt theorem therefore implies that the optimal inter-molt spacing is:

$$\Delta\phi^* = \frac{\tau_{\text{molt}}}{2} \propto \frac{1}{2\sqrt{\lambda}}$$

In rapidly changing environments (high $\lambda$), molts are frequent and spacing is tight — the fleet must carefully choreograph molts to avoid capability gaps. In stable environments (low $\lambda$), molts are rare and spacing is generous — the fleet can afford simultaneous molts with minimal risk.

7.4 Implementing Staggered Molts

Staggered molting requires heterogeneous molt triggers. Two mechanisms achieve this:

  1. Heterogeneous $C_i$ (P56, Corollary 5.1): Agents with different conservation constants naturally molt at different times, since their $\gamma_i$ reach the over-crystallization threshold at different rates.

  2. Artificial offset injection: The fleet coordinator can delay individual molts by up to $\tau_{\text{molt}}/2$ without significantly degrading the molting agent (the agent continues operating at $\gamma \approx 0.95$ during the delay, which is high but not yet critical).

The second mechanism connects to P03’s Open Problem 1 (“when should an agent molt?”). Our answer: an agent should molt when both (a) melt pressure exceeds crystallization capacity (P56’s answer), and (b) the fleet-wide molt schedule permits it (our additional constraint).

8. Experiments

8.1 Molt Function Calibration

Setup. Instrument 10 agents with shell-transition logging. Record capability metrics (latency, accuracy, skill availability) at 100ms intervals during 50 molts total. Fit $m(t)$ parameters ($\tau_{\text{molt}}$, $t_{\text{dip}}$, dip depth) via maximum likelihood.

Prediction. $\tau_{\text{molt}}$ ranges from 30s to 5min depending on shell complexity. Dip depth $m(t_{\text{dip}})$ is in $[0.15, 0.25]$. Dip timing $t_{\text{dip}} / \tau_{\text{molt}}$ is in $[0.15, 0.25]$.

Falsification. If the molt function is monotonic (no dip), Definition 1’s (M2) fails and Theorem 1’s CTF probabilities are overestimates.

8.2 CTF Rate Measurement

Setup. Deploy a 9-agent fleet under P47’s coordination protocols. Induce molts with known timing. Measure CTF events (task failure, degraded output detected by downstream validation, gossip contamination).

Prediction. CTF rates match Theorem 1’s predictions: $\approx 0.73$ for MS during master molt, $\approx 0.41$/cycle for CW, degradation radius $\approx 1.0$ for Peer.

Falsification. If CTF rate is below 0.1 for all patterns, the fifth failure mode is negligible in practice.

8.3 $\delta$-Based Detection Latency

Setup. Compute $\delta_i(t)$ at 1s intervals for 5 agents undergoing 20 molts total. Measure detection latency (time from molt onset to $\delta$ exceeding threshold $\delta_0 + 0.3$).

Prediction. Median detection latency $\leq 0.1\tau_{\text{molt}}$. False positive rate (detecting molt during non-molt $\delta$ fluctuations) $< 5\%$.

Falsification. If detection latency exceeds $0.3\tau_{\text{molt}}$, the detector is too slow for real-time scheduling correction.

8.4 Staggered vs. Simultaneous Molt Comparison

Setup. Two 9-agent fleets: one with staggered molts (spacing $\tau_{\text{molt}}/2$), one with simultaneous molts. Run 100 fleet-wide molt cycles. Measure CTF count, task completion rate, and end-to-end throughput.

Prediction. Staggered fleet has $\approx 9\times$ fewer CTF events and $\geq 2\times$ higher throughput during molt windows. Outside molt windows, both fleets perform identically.

Falsification. If staggered molts provide $< 2\times$ CTF reduction, the staggered molt theorem’s assumptions are violated (e.g., $T_{<0.5} > \tau_{\text{molt}}/2$).

9. Conclusion

This paper identifies and formalizes the fifth failure mode of multiagent coordination: capability transition failure during molting. While Paper 47’s protocols successfully eliminate deadlock, livelock, starvation, and cascade under the assumption of static capabilities, they are blind to the transient capability degradation that occurs when agents molt.

The four theorems establish a complete framework for molt-aware coordination:

  1. The Capability Transition Model (Definition 1) captures the non-monotonic recovery dynamics, including the critical molt dip that makes agents most vulnerable precisely when their self-reported status is most misleading.

  2. The Fifth Failure Mode Theorem (Theorem 1) quantifies CTF probability for each coordination pattern, showing that Master-Slave is most vulnerable ($\approx 73\%$ CTF during master molt) and Peer is least vulnerable (partial degradation via gossip).

  3. The Molt Detection Theorem (Theorem 2) leverages Paper 57’s conservation deviation $\delta$ as a real-time molt detector with sub-$0.1\tau_{\text{molt}}$ latency, enabling the fleet to detect molts without relying on agent self-reports.

  4. The Staggered Molt Theorem (Theorem 4) proves that spacing molts at $\tau_{\text{molt}}/2$ intervals reduces CTF rate by approximately $9\times$, connecting fleet-level coordination policy to individual agent dynamics from Paper 56.

The central message is that any coordination protocol assuming static agent capabilities is fundamentally incomplete. Molting is not a rare anomaly but a periodic necessity (P03, P56), and coordination systems must be designed to accommodate the capability transients it produces. The $\delta$-based detection and $\alpha_{\text{molt}}$-based scheduling we propose extend Paper 42 and Paper 57’s frameworks into the multiagent coordination domain, providing the first complete theoretical treatment of coordination under capability transition.

Limitations and Future Work

References

SuperInstance Fleet · served from the edge by Cloudflare Workers · fleet home