More Agents Is All You Need

• The paper demonstrates that deploying additional agents yields near-optimal utility through independent sampling and statistical ensembling.
• It shows that cost-speed trade-offs in multi-agent systems require balancing accelerated search outcomes with increased resource expenditure.
• The work illustrates that ensemble techniques in LLMs and structured agent roles can boost accuracy dramatically, enhancing reliability and production readiness.

The maxim "More Agents Is All You Need" encapsulates a family of results—spanning information search, delegated mechanism design, LLM ensembles, and real-world multi-agent orchestration—that demonstrate systematic performance gains when additional agents are deployed, often with surprising ubiquity and robustness. In many domains, the inclusion of more agents, even absent sophisticated cooperation or incentive adjustments, provides non-trivial improvements in efficiency, robustness, reliability, and solution quality through mechanisms such as ensembling, parallelization, and redundancy. However, the scope, modeling assumptions, and cost trade-offs underlying the effectiveness of many-agent architectures are nuanced, revealing both powerful positive results and sharp limitations.

1. Foundational Models: Delegated Search and Mechanism Design

Delegated search models formalize the scenario in which a principal seeks to maximize utility over a stochastic solution space $E$ but must delegate search to $n$ agents, each with (possibly misaligned) utility. The prototypical mechanism is a single-proposal threshold rule: each agent samples from their assigned subset $E_i\subseteq E$, then proposes their best outcome meeting a threshold $t$, and the principal selects the proposal with maximal utility or rejects all if no proposal meets $t$.

The principal’s achieved utility is compared to the first-best benchmark

$U^* = \mathbb{E}\Big[\max_{e\in E} x(V(e))\Big]$

and a mechanism achieves approximation ratio $\alpha(n)$ if the principal’s expected utility is always at least $\alpha(n) U^*$ (Bechtel et al., 2024).

A key theoretical advance is the sharp characterization of how $\alpha(n)$ scales with the number of agents. Define $p_n$ as the root in $n$0 of

$n$1

Then,

$n$2

and, even in adversarial settings,

$n$3

with both lower and upper bounds converging to $n$4 as $n$5.

Significantly, this asymptotic optimality emerges not because of direct competition—adversarial agents cannot collude to defeat the rate—but rather because the union of agent samples increases the probability that an outstanding proposal is found and submitted. This effect is robust to strategic or adversarial play, provided basic symmetry and independence assumptions hold (Bechtel et al., 2024).

In Bayesian and prior-independent mechanisms for multi-agent delegation—see also (Hajiaghayi et al., 2023)—the additive loss from using a prior-independent mechanism with $n$6 agents decays with $n$7, and for symmetric settings the guarantee on the principal’s utility approaches optimality as $n$8 grows. Explicitly, with $n$9 i.i.d. draws per agent, for symmetric Unif$E_i\subseteq E$0 distributions, expectations satisfy $E_i\subseteq E$1 as $E_i\subseteq E$2, so the additive gap vanishes (Hajiaghayi et al., 2023).

2. Collective Search: Speed-Quality-Cost Trade-offs

In distributed target search, increasing the number of agents reduces the expected time to hit the target, but also raises launch and sustainment costs. The optimal deployment policy is defined by the stochastic cost functional

$E_i\subseteq E$3

where $E_i\subseteq E$4 is first passage time, $E_i\subseteq E$5 is the number of agents actually deployed, and $E_i\subseteq E$6 is the total agent-time (Meyer et al., 2024).

Under log-convex single-agent survival $E_i\subseteq E$7, the cost-optimal launch policy exhibits regime behavior:

• If $E_i\subseteq E$8 is flat ("mild" convex), launch many agents simultaneously at $E_i\subseteq E$9 ($t$0).
• If $t$1 is steep, only one agent is launched at $t$2, with others staggered at optimal intervals.
• For exponential and algebraic tails in $t$3, the optimal spacing changes qualitatively.

Closed-form formulas determine the optimal number to launch at $t$4 and the cadence for further launches. Cost-minimizing strategies can include time-staggered launches or resort to stochastic resetting, with the superiority of "more agents" depending on launch, sustain, and reset cost regimes. When launch and sustain costs are low, increasing agent count always hastens search, but incurs higher resource expenditure; the dominance of more agents is then purely a question of application-specific cost structure (Meyer et al., 2024).

3. Sampling-and-Voting in LLMs: The Agent Forest Principle

The "Agent Forest" paradigm in LLMs postulates that simply increasing the number of independently instantiated agents, each issuing a solution, then aggregating via majority voting (or similarity-based scoring), steadily improves accuracy across diverse tasks (Li et al., 2024). For a query $t$5, $t$6 agents are each sampled (with prompt, temperature, etc.), providing outputs $t$7, which are then aggregated: $t$8 where $t$9 quantifies vote or answer similarity (e.g., exact match, BLEU score).

Extensive empirical evaluation demonstrates strictly monotonic accuracy gains from $t$0 up to $t$1; for instance, Llama2-70B on GSM8K jumps from $t$2 (single) to $t$3 (N=40) (Li et al., 2024). Gains are largest for more difficult tasks (e.g., a 200% relative gain for MATH vs 69% for GSM8K with Llama2-13B). Sampling-and-voting is orthogonal to, and stacks additively with, existing prompt engineering, chain-of-thought, and debate frameworks.

However, the returns diminish for very large $t$4 or extremely difficult tasks. The computational cost increases linearly in $t$5, and the mechanism does not always resolve error accumulation in pathological settings.

The persistent gain arises from statistical ensembling—by bootstrapping more samples, the aggregate answer converges to the most stable hypothesis, counteracting non-systematic errors of individual agents.

4. Production-Oriented Multi-Agent Coordination and Specialization

In high-stakes automation such as incident response, orchestrating distinct, specialized LLM agents dramatically enhances determinism and correctness. In MyAntFarm.ai, a "multi-agent copilot" comprises three sequential specialists (diagnosis, remediation, risk assessment), each with narrow objectives and hardened prompt templates (Drammeh, 19 Nov 2025).

Empirical results across 348 controlled trials reveal that multi-agent orchestration yields:

• Actionability rate: 100%
• Specificity improvement: 80x
• Correctness improvement: 140x
• Zero quality variance across trials

Average comprehension latency matches the single-agent baseline, revealing that the benefit is not speed but the deterministic transformation of output quality. These findings recast multi-agent orchestration as a minimal requirement for production readiness in LLM-driven decision support.

The Decision Quality (DQ) metric provides a composite, operationally meaningful evaluation targeting validity, specificity, and correctness. Multi-agent decomposition, by restricting each agent’s context and objectives, isolates faults and prevents deadlocks, facilitating both transparency and structured aggregation (Drammeh, 19 Nov 2025).

5. Organizational Models: Teams of Rivals and Layered Critique

Deploying agents with strict role boundaries, explicit division of planning, execution, and layered critique—an “AI Office”—approaches human organizational hierarchy. This architecture leverages not just consensus, but “rivalry”: critics hold absolute veto over outputs, orchestrators enforce correctness, and agents interact via fully decoupled code execution environments (Vijayaraghavan et al., 20 Jan 2026). The probability of user-facing error is sharply reduced:

$t$6

Compared to 60% baseline accuracy for single-agent approaches, multi-agent teams achieve 90% accuracy even in complex financial reconciliation tasks, with manageable overhead (38.6% token, 21.8% latency) and robust error interception (Vijayaraghavan et al., 20 Jan 2026). Modularity of agents enables frictionless expansion, upgrade, or specialization, avoiding regression in existing pipelines.

A nuanced insight is that the gain is due to the deliberate orchestration of role-diverse, opposing-incentive agents—especially critics—not merely increased headcount.

6. Adaptive and Task-Aware Agent Deployment

Static addition of agents is not always optimal: adaptive design paradigms dispatch more agents only on detection of task complexity, achieving cost–benefit trade-offs. In LLM-based code debugging, a Main Agent spawns specialized sub-agents as dictated by bug type and complexity score:

• Simple (syntax-only): 1 agent
• Medium (logic/reference): 2–3 agents
• High (multi-error/algorithmic): 3–5 agents

Experimental results show +6% (GPT-4) to +18% (Llama3) fix-rate improvement over single prompt baselines for complex cases, with agent count and iterative depth scaling linearly with problem complexity. The adaptive agentic design constrains focus drift and resource expenditure, mitigating the drawbacks of static multi-agent deployments (Majdoub et al., 25 Apr 2025).

7. Limits, Adversarial Robustness, and Orchestration Theory

Although ensembling and multi-agent approaches yield consistent gains in clean and moderately perturbed settings, robust adversarial performance plateaus when input errors are highly correlated or semantic, as majority voting fails to rectify systematic misinterpretations. For instance, in mathematical reasoning under adversarial typo attacks, sampling-and-voting boosts accuracy from $t$7 (N=1) to $t$8 (N=25) under clean conditions but the gap to perturbed accuracy remains nonzero for real-world typo attacks even as $t$9 (Alavi et al., 10 Nov 2025).

Orchestration theory formalizes when more agents actually help. If all agents have identical performance or cost structures across all tasks/regions, orchestration provides no benefit:

$U^* = \mathbb{E}\Big[\max_{e\in E} x(V(e))\Big]$0

But when agent strengths cross by region, dynamic allocation (orchestration) produces strict gains (Bhatt et al., 17 Mar 2025). Empirical and theoretical analyses confirm that "more agents" only justifies its cost when capitalizing on heterogeneity or specialization.

8. Synthesis and Open Problems

The "More Agents Is All You Need" principle is robust across theoretical delegation, search, LLM ensembles, and organizational AI, provided agents are independent (in sampling or information), or bring heterogeneity or role specialization. The mechanism of improvement is not mere competition but larger aggregate sample sets, diversity, or modularity of roles.

Significant open problems include formal quantification of diminishing returns, optimal dynamic agent allocation given real-time task difficulty, closing the gap between lower and upper bounds in mechanism design, and architecting ensemble methods robust to adversarial correlations. Extensions to combinatorial settings, richer strategic spaces, and cost-sensitive orchestration remain active research areas.

In conclusion, with appropriately controlled costs, modeling assumptions, and orchestration logic, increasing the number of agents systematically amplifies capability, reliability, and robustness across a spectrum of multi-agent AI tasks and mechanisms (Bechtel et al., 2024, Hajiaghayi et al., 2023, Li et al., 2024, Drammeh, 19 Nov 2025, Vijayaraghavan et al., 20 Jan 2026, Majdoub et al., 25 Apr 2025, Alavi et al., 10 Nov 2025, Meyer et al., 2024, Bhatt et al., 17 Mar 2025). However, the exact benefits, limitations, and applicability must be evaluated within the specific structure and practical constraints of each domain.