agent

1、Background: The release of Tencent’s Hunyuan Hy3 sends a clear signal that domestic large-model competition is shifting from “parameter-scale showdowns” to “business usability and Agent capability deployment.” This model uses a Mixture-of-Experts (MoE) architecture with 295B total parameters and 21B activated parameters. The core idea is to maintain strong capability while reducing inference cost. Compared with simply chasing larger parameter counts, MoE emphasizes calling specialist networks on demand, which makes it easier to balance performance, cost, and response speed in high-frequency business scenarios. Hy3 has now been integrated into the WeChat ecosystem to serve an extremely large user base, meaning it is not just a lab model, but infrastructure-level capability entering a real product environment.

2、Analysis: Based on disclosed information, Hy3 is positioned as an Agent for LLMs. Its focus is not merely on “chatting,” but on “getting tasks done.” The internal WorkBuddy task success rate increased from 72% to 90%, while time was reduced by 34%. This suggests Tencent cares more about end-to-end execution efficiency, including task decomposition, tool calling, result verification, and feedback-driven correction. It performs strongly in coding, office work, and complex task planning—areas that align with enterprise-level AI needs. Examples include generating web pages, creating PPTs, handling office workflows, and assisting R&D—application directions that can directly boost productivity. Notably, the model has the ability to self-check and proactively explain shortcomings, which is crucial for reducing hallucinations and improving controllability. However, pure visual capability is still seen as a weakness, indicating that multimodal abilities are not comprehensively leading. In the future, further strengthening will be needed in image understanding, video reasoning, and cross-modal tasks.

3、Impact: For the AI industry, Hy3’s significance lies in validating the path of combining “model capability” with a “super app entry point.” WeChat has a massive user base and abundant scenarios. If AI capabilities are deeply integrated into chat, search, office, content generation, mini programs, and enterprise services, it will accelerate the transition of Agents from concept to everyday use. For enterprise customers, improved model efficiency may mean lower calling costs and stronger willingness to deploy, pushing AI from pilots to large-scale applications. For the market, this will also intensify competition among domestic large-model vendors in ecosystem building, compute resources, data, and product experience. The true watershed ahead is not just leaderboard performance, but who can continuously reduce error rates and improve task completion in real business settings, and form a sustainable commercial closed loop. Overall, Hy3 represents large-model applications entering a “pragmatic phase”: less talk about concepts, more focus on deployment; less obsession with parameters, more on efficiency and results. 🚀

#AI #腾讯混元 #Agent

$GROK JUST OUTPERFORMED CLAUDE AND OPUS ON COST AND PERFORMANCE 🔥

Grok 4.5 just dropped a 51.4% on AutomationBench-AA — ahead of Claude and Opus — and did it at $0.34 per task against their $1.35+. That's a massive efficiency gap for AI agents.

The model is already live on SpaceXAI's V9 foundation and the benchmark validates what the team pitched. Independent data like this usually brings attention to tokens tied to real infrastructure. Volume could shift fast if traders start pricing in the adoption story.

Are you watching $GROK ride this catalyst?

Not financial advice. Always manage your risk.

#GROK #AI #Benchmark #Agent

🔥

SpaceX’s Cursor on iOS: the cloud proxy stays resident on your phone, and remote control of your computer is no problem either. The “AI Agent” story keeps adding fuel—though everyone keeps shouting about disruption, only a few have truly landed in a form you can call on anytime. Is this a real need or just skin-deep automation? Let’s keep an eye on it first; a small bet for fun. #AI #Agent $FET $OLAS
$WLD

GUYS I was watching my recursive AI #AGENT run yesterday and noticed a tiny, ann0ying pause right before it generated each response.

At first, I assumed the model itself was just slow.

But as my agent started executing complex, multi-step workflows, those milliseconds began adding up.

I realized the real performance bottleneck is not the GPU .....

the AI model's computation speed....

It's the constant cryptographicc signature validations needed to approve and pay for every single reasoning step......

For me, this creates what I call a "Sign-to-Think Ratio."

If an AI spends more time signing transacti0ns to prove it can run than it does actually thinking, the system chokes....

This is why @OpenGradient integration of Permit2 on Base is a game-changer.

By batching token approvals, it prevents transaction spam from draining the agent's verification budget.

I tested this lowlatency setup myself at chat.opengradient.ai....

and it feels as seamless as a #centralized app, but with complete hardware enforced privacy under the hood.....

Personally, I'm buying credits to run my developer workflows....

I think we are f0cusing way t00 much on buying faster chips when we should be optimizing the math that validates them.

Do you think signature congestion is the biggest roadblock for on chain AI?

#OPG $OPG #DeAI $TAC $GWEI

1、Background

Today Hermes Agent launched the MoA (Mixture of Agents, mixed intelligence) feature. This is an important upgrade to the product form of the open-source agent platform. Previously, instead of treating multi-model collaboration as a bottom-layer tool, MoA is now packaged as a “virtual model provider.” Users can call it directly just like they switch between ordinary large language models. This design significantly lowers the barrier to use, and it also means that multi-model coordination is shifting from a developer-oriented play style to a standard capability entry point for everyday users 🤖

Based on the disclosed information, users can choose MoA directly via /model, or make a one-off call via /moa [prompt]. This suggests that Hermes Agent is not just adding a button—it is actively trying, at the interaction layer, to turn “multi-model orchestration” into an immediately usable product experience.

2、Core Analysis

The key value of MoA lies in turning the combination of a “reference model + aggregator model” into a more stable inference pipeline. The reference model first generates analysis opinions based on a simplified version of the conversation text. Then, the aggregator model combines the full system prompt, tool schema, and context to produce the final response and execute any tool calls. Put simply: one model is responsible for “thinking,” and another model is responsible for “doing.”

There are two aspects of this mechanism that are worth paying attention to. First, the reference model does not access the complete tool history, which reduces context-noise interference and helps produce more focused analysis output. Second, the aggregator model retains final decision-making authority. Under complete rules, it can unify task execution, reducing style conflicts and action mismatches that can arise when multiple models output in parallel directly.

If, as shown in the article, HermesBench’s subsequent test results allow MoA to outperform a single model on the leaderboard, then its significance is not only that it achieves a higher score. More importantly, it validates a trend: in future agent competition, the edge will not come solely from single base-model parameters and raw capabilities, but from “model collaboration architecture” and “task orchestration efficiency.”

3、Industry Impact

For the AI Agent track, this development may push product competition into a new phase. The market has recently been broadly focused on the upper limits of model capability, but Hermes Agent’s moves indicate that engineering and organization can also create incremental value. For developers, MoA makes multi-model collaboration easier to deploy; for users, it may bring higher-quality, more stable execution experiences for complex tasks.

From a Web3 perspective, if the open-source agent platform can productize MoA first, there will be room for expansion in scenarios such as on-chain data analysis, investment research support, automated customer service, and content generation. Especially in environments where costs and outcomes need to be balanced, splitting work across multiple models may offer better cost-effectiveness than forcing a single model to handle everything.

4、Summary

Overall, today’s launch of MoA in Hermes Agent is not just a feature update—it feels more like a forward-looking exploration of an agent product paradigm. The signal it releases is very clear: the next stage of AI Agents may shift focus from “stronger single models” to “better collaborative systems.” If subsequent benchmark testing and real user feedback continue to validate the results, MoA is likely to become one of the important standard configurations for agent platforms 📈

#AI #Agent #crypto

$AGENT STANDARD LAUNCHED — NEW FRAMEWORK FOR AI INTERCONNECTION 🔥

The newly released national standard series for AI intelligent agent interconnection covers seven key areas including architecture, identity, discovery, and interaction. This creates a closed-loop system for cross-domain agent collaboration.

Standardized identity authentication and end-to-end traceability reduce custom development time and strengthen trust in automated agent interactions. As these standards roll out, expect faster deployment and broader adoption of AI agent ecosystems.

Are you watching which projects align first with this new framework?

Not financial advice. Always manage your risk.

#AGENT #AI #Standards #CryptoAI #TechAdoption

🔥

Doubao launches a professional version starting at 68 CNY, with the debut supporting the Office Agent mode for PC operations.

Doubao, under ByteDance, has rolled out a professional version starting at 68 CNY, which supports users in directly controlling their computers to complete office tasks through the AI Agent. This includes file organization, data processing, application operations, and more, effectively upgrading AI from just a chat tool to a true office assistant.

Why it matters: The evolution of the AI Agent from conversational interaction to direct computer control marks a crucial shift in AI office products from being auxiliary tools to autonomous executors.

#AI #人工智能 #豆包 #Agent

【Dev Reminder】The AI Agent infrastructure B.AI will be deprecating the old API Key on June 22.

If you're still trading with the old Key, you need to act fast:
1) Generate a new API Key
2) Swap out the old credentials in your current setup
3) Run your tests to ensure the call chain is smooth

These updates might seem like just a key switch, but for projects relying on Agent services, automated workflows, or backend calls, failing to migrate could lead to service interruptions. It's best not to wait until the last minute, especially in production, where you should allow time for rollbacks and troubleshooting.

#AI #Agent #Dev Tools

[Developer Reminder] B.AI's AI Agent infrastructure will deprecate the old API Key on June 22nd. If your application, automated processes, or third-party integrations are still using the old Key, it is recommended to generate the new API Key in advance and complete the configuration replacement and connectivity testing.

This type of update may seem like just a "key migration," but for Agent services that rely on the API, delayed processing can directly lead to call failures, task interruptions, or business anomalies. Teams already integrated with B.AI should include this migration in their operations and maintenance checklist to avoid the risks associated with concentrated processing close to the deadline.

#AI #Agent #开发者 ԥ

Visa has issued cards for AI agents. Alchemy's AgentCard is now connected to the Visa network, allowing AI to spend directly.
This isn't some UFO narrative; it's a legit payment channel that's been opened up. Buying computational power, paying for APIs, and settling data fees are all fully automated, making it feel like on-chain Agents are starting to exhibit 'wallet behavior'.
We're still in the infrastructural phase, waiting for the first consumer-grade Agent to generate data. Once that happens, this AI payment narrative will be more than just hype; keep your eyes on this lane. #AI #PayFi #Agent $BTC

Let AI manage the funds, which sounds fancy as autonomy, but in reality, it’s just giving robots a paycheck. Next up, it’ll definitely take the private keys and start trading crypto on its own. The narrative of on-chain Agents has just begun; just wait for an AI whale to pump the market. #AI #Agent $FET

Y Combinator drops a game-changing AI Agent: Just shoot a text to launch and run a full-on business

The well-known incubator Y Combinator has rolled out the "Locus Founder" AI Agent. Users can just send a text via iMessage, SMS, or Telegram describing their business idea, and the AI will handle everything from business formation to operations and USDC payment settlements. From product ideation to actual operations, it’s all taken care of by the AI itself.

Why it matters: This is the most radical attempt at deploying AI Agents in the business realm—AI is no longer just an "assist tool"; it’s stepping up as an independent business operator, set to completely revolutionize the startup landscape and Web3 payment scenarios.

#YC #AI #Agent #Web3

Hermes Agent launches async sub-agent and Stripe payment skills

Nous Research announces major updates to the Hermes Agent framework: the async sub-agent backend task feature allows users to chat normally while the sub-agent is running, with the main window no longer locking up; plus, three Stripe payment integration skills can be installed and used directly via the hermes skills install command.

Why it matters: The async sub-agent makes multitasking possible, and combined with payment skill integration, the AI Agent is evolving from a chat tool into a self-sustaining system capable of executing business operations.

#HermesAgent #AI #Agent #Web3

[Data Stream] AI Agent Tokens: What Can On-Chain Data Tell Us?

Recently, AI Agent sector tokens have been showing strong performance, but FOMO is also on the rise. Let’s do a ‘check-up’ on this sector using on-chain data.

📊 Key Metrics Overview:

1️⃣ Holding Concentration (Top 10 Addresses' Holding Ratio)
The higher the token holding concentration, the lower the pump cost, but the greater the selling pressure risk. Most tokens in the AI Agent sector have a top 10 holding ratio in the 30-60% range, which is considered moderate concentration.

2️⃣ Contract Interaction Activity
The contract call volume of the last 30 days vs. the 90-day average: mainstream AI tokens have increased by 2-5 times, indicating genuine usage is on the rise and it's not just speculation.

3️⃣ Whale Position Changes
By tracking known whale addresses through on-chain tags, the net buys in the last week have concentrated on $FET , $GRASS , and other established AI tokens, while newer narrative tokens have seen net outflows.

4️⃣ Liquidity Coverage Ratio
CEX deposit address balance / average daily trading volume, the higher the ratio, the stronger the cash-out ability. A healthy range is > 3x.

🔍 Conclusion:
The sector is generally hot, but there’s internal structural differentiation—established AI tokens have genuine on-chain data support, while newer narrative tokens are more about capital rotation effects.

Trading Advice: Be cautious with chasing highs; focus on assets with consistently increasing on-chain activity rather than just hype based on concepts.

#AI #Agent #链上数据 #CryptoInvestment

📰 Crypto Market Highlights

1. OpenRouter Launches Fusion Composite Model Interface
OpenRouter recently rolled out the Fusion composite model solution, allowing the same prompt to be distributed in parallel to multiple large models. The final answers are then integrated through a referee and composite model. The latest benchmark tests show that multi-model collaboration significantly outperforms traditional single models in complex reasoning and deep research tasks, showcasing the value of 'multi-perspective complementarity'. The market's focus is on the potential for this solution to achieve results close to top-tier closed-source models at a lower cost, accelerating the evolution of AI infrastructure towards 'model orchestration + result synthesis'.

2. Multi-Model Mixing Boosts Cost-Effectiveness as an Industry Highlight
According to public test results, combinations of models from different vendors perform stronger in complex tasks, enhancing both answer stability and reasoning coverage. Notably, even dual-channel collaboration and self-synthesis with the same model showed a significant score improvement. This indicates that composite reasoning is shifting from 'stacking parameters' to 'rearranging', which may lead to increased market attention on reasoning layers, middleware layers, and AI service aggregation platforms. The relevant technological pathways are worth ongoing tracking.

3. Databricks Opens Up Omnigent for Agent Management
Recently, Databricks has open-sourced the meta-assembly framework Omnigent, which supports operation on multiple existing agent tools and transforms intelligent agents under different frameworks into interoperable components, alleviating issues of fragmented interfaces and collaboration difficulties. Its core highlight is the ability to execute stateful security policies at the meta-assembly layer, no longer relying solely on prompt constraints. For enterprise-level AI deployment, such infrastructure with cross-agent orchestration, permission review, and process intervention capabilities is becoming a key support for deploying intelligent agent systems.

4. Security Approvals and Cost Control Become Key Focus for Agent Deployment
Omnigent has also strengthened real-time risk control, budget management, and collaboration capabilities. For example, it can interrupt processes and request manual approval upon detecting high-risk actions, and can automatically pause tasks when model invocation costs reach their limits. The system also offers sandboxed network request interception and team-shared session functionality, highlighting that agent applications are shifting from 'can tasks be completed' to 'are they controllable, auditable, and collaborative'. This trend may drive enterprises to place greater emphasis on AI governance, cost monitoring, and compliance infrastructure development.

#AI #Agent #crypto

Cambridge & Chicago Universities Open Source DecentMem: Decentralized Memory Boosts Multi-Agent Collaboration Efficiency by 24%

The teams from Cambridge and Chicago Universities have open-sourced the multi-agent memory framework DecentMem, using decentralized private memory to replace traditional global shared memory. Research shows that shared memory leads agents to converge on similar decision paths, while DecentMem maintains cognitive diversity by preserving each agent's private memory. In tests with AutoGen, DyLAN, and AgentNet, DecentMem averaged an 8.6% improvement over centralized memory baselines, with a peak enhancement of 23.8%, while halving token consumption.

Why it matters: DecentMem addresses the core issue of "division of labor failure" in multi-agent systems at the architectural level, paving the way for a more efficient AI agent collaboration network.

#AI #多智能体 #开源 #Agent

Databricks has rolled out the open-source Agent orchestration tool, Omnigent, tackling the challenges of multi-Agent collaboration and security control.

Databricks has open-sourced the Agent orchestration framework Omnigent under the Apache 2.0 license, which runs on existing tools like Claude Code, Codex, and Pi, allowing different frameworks' agents to be transformed into interoperable system components. Omnigent implements stateful security controls directly at the orchestration layer, supporting the interception of git push actions after agents download npm dependencies, requesting manual approval, or setting LLM cost limits to pause operations when cumulative costs hit $100. The framework also integrates a network request sandbox to prevent sensitive information leaks.

Why it matters: Omnigent fills the interoperability gap in multi-Agent orchestration, providing crucial security control infrastructure for AI Agents transitioning from experimental to enterprise-level deployment.

#Databricks #AI #Agent #open-source

📰 Crypto Market Hotspot Dispatch

1. Nvidia's Blackwell Sets New Benchmark for AI Hardware Efficiency
The latest benchmark, aa-agentperf, shows Nvidia's Blackwell significantly outpacing competitors in AI workload scenarios. The tests replay real programming trajectories and use the number of concurrent agents supported per megawatt of power consumption as a key metric. Results indicate that the GB300 NVL72 can handle about 61,400 concurrent agents under the same power budget, representing a more than 20-fold improvement over the H200, with substantial enhancements in single-card concurrency as well. This suggests that the infrastructure costs for high-concurrency scenarios like AI agents, automated programming, and customer service are likely to continue dropping, accelerating the competition in computational efficiency.

2. AI Infrastructure Competition Heats Up, AMD Faces Greater Performance Pressure
From the results of this AI hardware testing, market focus has shifted from purely training performance to inference efficiency, concurrency capacity, and energy output per unit. Nvidia's Blackwell, with its liquid-cooled rack system and high-density deployment capabilities, has established a stronger edge in AI application scenarios, putting pressure on competitors like AMD. For the crypto market, the rising heat in the AI computing supply chain could continue to impact the sentiment pricing of GPUs, data centers, power resources, and AI concept assets, with funds increasingly focused on the new narrative of "efficient inference."

3. OpenRouter Tests Subagent Tool to Propel Multi-Model Collaboration
OpenRouter has recently launched the server-side proxy tool openrouter:subagent, enabling the main model to dispatch specific sub-tasks to smaller, lower-cost models during the generation process, which then return results. This mechanism helps compress invocation costs while maintaining overall quality and enhancing flexibility in executing complex tasks. If the working model integrates search, scraping, and other tools, it can first complete retrieval and multi-step inference before feeding back to the main model, showcasing how AI applications are transitioning from "single model responses" to "multi-agent collaboration."

4. Subagent Architecture Enhances Practicality, but Context Management Remains Key
It’s important to note that the subagent solution is not fully automated. The working model cannot directly read the main model's context, so the main model must provide complete background information in the task description; otherwise, execution quality may be affected. To prevent infinite recursion and resource control issues, OpenRouter has also implemented safeguards against self-reference, limited nesting depth, and overall task count caps. Overall, these tools are more suitable for developers and enterprise workflows, and may accelerate the deployment of low-cost AI agent products, further enhancing market attention on the Agent sector.

#AI #Agent #Nvidia

OpenRouter drops the subagent tool: Big models can delegate side tasks to smaller models mid-generation.

OpenRouter launched the server-side proxy tool openrouter:subagent, allowing big models to hand off independent side tasks to smaller, cheaper candidate models while generating content. The results of these side tasks come back as outcomes that the main model integrates. The working model can also be equipped with online search, web scraping, and other standalone tools for multi-step reasoning in a sandbox environment. To prevent infinite recursion, OpenRouter has introduced nested depth limits and hard caps.

Why it matters: subagent pioneers a new paradigm for task collaboration between models, significantly lowering the reasoning costs for complex Agent tasks.

#AI #OpenRouter #人工智能 #Agent