Hybrid Agents Break Through: 80% Cost Reduction, AI Autonomously Beats Minecraft

AI agents are moving from "demos" to "delivery." A recent wave of research shows hybrid reasoning architectures compressing inference costs to a new order of magnitude, while multi-agent collaboration can now complete long-horizon, highly complex tasks like Minecraft with no human intervention. However, when the evaluation scenario shifts to real-world customer delivery, the agents' shortcomings are equally exposed.

Hybrid Reasoning Architecture: Teaching Agents to "Think More, Say Less"

Dynamically Switching Between Latent Space and Text

Traditional agents must output full text at every reasoning step, consuming large amounts of tokens and responding slowly. The newly proposed HybridAgent introduces a dual-mode mechanism: it enters latent reasoning when deep thought is needed, and switches back to text mode when interacting with tools or the environment. An intermediate HybridLink component handles seamless transitions between the two modes while preserving tool-calling capability.

The core insight of this design is that not every reasoning step needs to be "spoken out loud." A large amount of intermediate deliberation can be completed in hidden vector space, externalizing only the key conclusions as text—thereby significantly reducing overhead while maintaining interpretability.

Latent reasoning refers to a mechanism that completes intermediate reasoning in the model's hidden vector space: instead of generating discrete text tokens at every step, the model directly uses hidden states as the "carrier of thought" for computation and transmission. Unlike mainstream Chain-of-Thought, it eliminates the encoding-decoding overhead of "translating thoughts into natural language" and then "reading the text back to continue reasoning," so it typically has advantages in token consumption and latency. The cost is that the intermediate process is not easily readable by humans, and interpretability and controllability decrease. HybridAgent's dual-mode approach attempts to balance both: staying in latent space when deep deliberation is needed, and returning to text when interacting with tools or the environment, so that key conclusions are externalized.

10.4% Average Gain Across 11 Benchmarks, Up to 80% Cost Savings

The experimental data is quite convincing. According to the arXiv paper (2610.18861), HybridAgent achieves an average improvement of 10.4% across 11 benchmarks spanning math, science, medicine, search, and agent reasoning, with token consumption reduced by 35.5% to 54.7%, inference speed improved by 1.8x to 2.6x, and cost savings of 2.7x to 3.7x. This benefit trend holds consistently across three different model scales, indicating a degree of scale robustness.

For practical deployment, this means the same budget can support longer task chains, or serve more concurrent requests at the same latency.

Four Agents, 2,783 Programs, Minecraft Conquered

End-to-End Completion with Zero Human Intervention

Another study set its sights on a more "wild" testing ground. Four Gemini 4 Argon agents entered a sandboxed Minecraft survival world with a single goal: defeat the Ender Dragon. Each agent controlled a player character through Mineflayer, coordinating via in-game chat and shared memory. No one assigned roles in advance, and no one wrote game strategies—all programs were planned, written, and debugged by the agents themselves.

The research team only monitored through a real-time dashboard, with no intervention throughout.

605 Deaths for a 27-Hour Victory

The cost was real. According to the arXiv paper (2610.22319), this agent team wrote a total of 2,783 programs and, after 605 deaths, completed the run in 27 hours and 14 minutes. This is the first end-to-end Minecraft completion record achieved autonomously by AI agents using only standard Mineflayer.

605 deaths are nothing to be ashamed of—they precisely demonstrate the agents' ability to recover from failure and iterate on strategies. In an open world, errors are the norm; the key is to keep self-correcting.

Mineflayer is a widely used Minecraft client protocol library in the Node.js ecosystem. It wraps the game's low-level network packets into programmable high-level APIs, allowing developers to locate, mine, craft, interact with entities, and even write automation scripts. In agent research, Mineflayer's significance lies in providing a stable, standardized, and reproducible access layer: agents don't need to learn low-level controls of "how to operate the game"—they only need to call APIs to execute actions. This is precisely why "using only standard Mineflayer" is emphasized as a constraint—it rules out the possibility of hardcoding low-level behaviors for specific tasks, making the completion result more reflective of the agents' own planning and programming capabilities.

MLExpert: Letting Models Do Machine Learning Engineering Themselves

A 35-Billion-Parameter MoE Agent

MLExpert is a 35-billion-parameter Mixture-of-Experts (MoE) agent model purpose-built for autonomous machine learning engineering. According to the arXiv paper (2610.22344), it employs hierarchical agent supervised fine-tuning (SFT) combined with windowed interactive reinforcement learning to handle long-horizon experimental iteration.

Mixture of Experts (MoE) is a sparsely activated model architecture: the network contains multiple parallel "expert" subnetworks, and a router selectively activates only a few experts for each token or input. Compared to dense models of the same parameter count, MoE uses only a subset of parameters during training and inference, thereby expanding model capacity while controlling computational cost. This "large capacity, sparse activation" property is especially important for agent models, because agents need to make multi-step decisions repeatedly over long-horizon tasks, multiplying the cost of each call. Note that MoE saves computation rather than memory footprint—all experts still need to be resident or distributed via parallelism.

MLE-Bench Lite Medal Rate Doubled

On MLE-Bench Lite, MLExpert raised the base model's average medal rate from 37.9% to 72.7%, matching larger frontier models such as Kimi-K3 and GLM-5.3. More importantly, this learned optimization capability generalizes to NatureBench Lite and algorithm optimization tasks, showing it is not overfitting to a single benchmark.

Real-World Delivery: FDE-Gym Pours Cold Water

95 Customer Tasks, Best Score 0.56

If the earlier progress seems encouraging, FDE-Gym's evaluation results provide a necessary reality check. The benchmark covers 95 end-to-end delivery tasks across 18 industries and 20 customer scenarios, each adapted by industry experts from real projects. Agents must clarify ambiguous requirements and discover unstated constraints in running systems.

The results are not promising. According to the arXiv paper (2610.22350), the best configuration achieved an average FDE-Gym score of only 0.56, solving just 14 of the 95 tasks.

FDE (Forward Deployed Engineer) is a role originating from enterprise software delivery practice: engineers are stationed at customer sites, responsible for understanding requirements in real systems, clarifying ambiguity, identifying constraints, and delivering usable solutions. FDE-Gym is named after this, meaning it assesses not the ability to "solve a clearly specified problem," but whether an agent can play this on-site delivery role—actively probing and delivering in environments where requirements are vague, systems are running, and constraints are unstated. This is also the fundamental difference from benchmarks like MLE-Bench Lite that target explicit metrics: the latter's success criterion is a score, while the former's is real business value generated in customer systems.

Unstated Constraints Are the Biggest Bottleneck

The research points out that discovering implicit constraints is the primary obstacle. Even on 23 tasks where "all key facts have been obtained," the best configuration achieved meaningful business improvements on only 10. This suggests the problem lies not just in information gathering, but in translating information into actionable solutions.

From lab benchmarks to customer sites, agents still have quite a long way to go.

DeepMind SIMA 2: Learning by Playing in Virtual Worlds

Google DeepMind released SIMA 2, an agent that can play, reason, and learn from users in virtual 3D worlds. According to DeepMind's official blog, SIMA 2 can follow instructions and complete tasks in simulated environments without game-specific integration. This offers another path for training general embodied agents: using virtual worlds as a scalable learning ground.

Two Signals in the Same Week

Viewed together, this body of research reveals a clear divide. On one hand, hybrid reasoning, multi-agent collaboration, and specialized agent models are driving costs down and capabilities up; on the other hand, FDE-Gym reminds us that real-world ambiguity and implicit constraints remain the agents' hardest challenge. Falling costs will accelerate deployment, and deployment will expose more real-world problems—these two forces will jointly define the focus of the next stage of agent competition.

FAQ

What is a hybrid reasoning agent?

It switches between internal latent-space reasoning and external text interaction, executing actions through tools while keeping the thought process inspectable.

How did the Gemini agents beat Minecraft?

Four agents wrote 2,783 programs, coordinated via chat, and defeated the Ender Dragon in 27 hours and 14 minutes after 605 deaths.

How much did MLExpert improve MLE performance?

It raised the average medal rate on MLE-Bench Lite from 37.9% to 72.7%.

What does FDE-Gym measure?

It measures AI agents' ability to deliver usable solutions for real customers across 95 tasks in 18 industries.

One-line Summary

Breakthroughs in both agent cost and capability, but real-world customer delivery remains the weak spot.

Paragraph Summary

This article reviews several key advances in recent AI agent research. HybridAgent dynamically switches between latent-space and text reasoning, achieving an average 10.4% improvement across 11 benchmarks, reducing token consumption by 35.5% to 54.7%, and saving 2.7x to 3.7x in cost. Four Gemini 4 Argon agents completed Minecraft in 27 hours after writing 2,783 programs and dying 605 times, with zero human intervention. The 35-billion-parameter MoE agent MLExpert raised the average medal rate on MLE-Bench Lite from 37.9% to 72.7%. However, the FDE-Gym benchmark poured cold water: across 95 real customer delivery tasks in 18 industries, the best configuration scored only 0.56 and solved 14 tasks, with unstated implicit constraints as the biggest bottleneck. The article notes that falling costs and real-world scenario exposure will jointly define the focus of the next stage of agent competition.