Solving one problem is good. Building a tool that solves a thousand is better.

With Hyra 2.0, Tencent shifts its AI research agent from solving individual problems to developing reusable scientific models, verification methods, and research tools.

Solving a scientific problem is one thing. Building a tool capable of solving hundreds of others is something else entirely. That's the shift Tencent is pursuing with Hyra 2.0, the latest version of its automated research system. After developing an agent capable of exploring hypotheses, running experiments, and improving its results, the Hunyuan team now wants to give it a more ambitious assignment: creating the models, software, and methods that will support future discoveries. The goal is to turn isolated AI achievements into lasting research capabilities.

This evolution builds on Hyra 1.0, which LHC covered in July. The original system relied on an experimentation loop that could propose solutions, execute them, and refine its approach based on evaluation results. Its performance claims came with a revealing warning: some agents had found ways to exploit their own tests, achieving higher scores without actually solving the problems. Hyra 2.0 does more than revisit those results. It attempts to move beyond research driven primarily by benchmark optimization, making the creation of reusable tools its central objective. Why Solving One Problem Is No Longer Enough

In its introduction to Hyra 2.0, Tencent distinguishes between two approaches to advancing automated research. The first involves assigning a problem to an agent, generating multiple attempts, and keeping the best solution. This approach has already produced results across several disciplines, including mathematics, computer science, and scientific simulation. But each success remains largely tied to the specific problem that produced it.

The second approach focuses on improving the instruments used to solve problems. A new molecular prediction model could support thousands of experiments. A better mathematical verifier could be used across numerous proofs. A more efficient data preparation method could accelerate multiple training projects. The initial investment may be greater, but its benefits can extend across successive research efforts.

Tencent identifies two broad categories of tools: those built around explicit methods, such as programs, rules, and research workflows, and those whose capabilities are acquired through training, including language models and scientific prediction systems. Hyra 2.0 is intended to contribute to both categories, moving beyond the direct search for solutions.

The ambition also connects to recursive self-improvement. If an agent helps build a better model, that model could subsequently enable the agent to develop more capable tools. But this feedback loop remains a research objective, not evidence that a system can improve indefinitely without human intervention. A Multiagent Organization Built to Work for Weeks

The most tangible change appears in Hyra 2.0's architecture. Complex scientific experiments cannot always be completed within a single session. Training a model, waiting for results, analyzing failures, revising a hypothesis, or coordinating multiple experiments can take days or even weeks. Tencent has therefore developed a three-level organizational structure: Overall, Mission, and Worker.

Overall provides general coordination. It communicates with human researchers, establishes priorities, and oversees available resources. Each Mission takes responsibility for a specific scientific objective, defines a research strategy, and determines which experiments should be conducted. Workers handle the practical tasks: reading code, modifying implementations, launching training runs, studying results, and proposing new hypotheses.

This structure allows multiple activities to proceed simultaneously. One Worker can investigate an alternative approach while another waits for an experiment to finish. A third can independently verify a conclusion. Tasks are stored in a persistent queue, together with their dependencies and priorities, allowing work to resume after interruptions or human approvals.

Hyra 2.0 is also integrated into Tencent's actual research environment. Agents can interact with researchers through WeCom, work in code repositories, and use computing platforms. Operations involving substantial computing resources or certain shared changes still require human approval. The autonomy Tencent is pursuing does not mean unrestricted access or decision-making authority. Agents That Verify Other Agents' Work

Verification plays a particularly important role in the new architecture. The problems observed with Hyra 1.0 demonstrated that an apparent improvement measured by an evaluator could sometimes conceal a methodological error or an unintended exploitation of the test's rules. With Hyra 2.0, Tencent introduces mechanisms intended to better distinguish a promising result from a conclusion supported by actual evidence.

When a Worker finishes a task, it must provide verifiable materials, including code versions, experimental results, logs, or execution traces. The Mission responsible for the project can examine these records and determine whether the observations genuinely support the reported conclusion. When independent verification is required, a separate Worker can revisit the original problem and supporting evidence without first being influenced by the executing agent's interpretation.

This addresses a familiar issue in multiagent systems: several agents can reinforce the same mistake when they share assumptions or repeat one another's conclusions. Separating certain verification stages may reduce that risk, although it does not guarantee independence equivalent to external scientific validation.

Tencent has also introduced a tracking environment called Research Board, where hypotheses, observations, experiments, and important decisions are recorded. Information remains accessible throughout a project's lifetime rather than depending entirely on the immediate memory of a conversation. A Shared Memory to Avoid Repeating the Same Experiments

Another major development concerns the preservation of