Meta acquired @ManusAI. Not a model company, they acquired an environment company, and the distinction is important. I have a solid argument favoring that intelligence cannot exist in isolation. It cannot be dissociated from the context and environment in which it operationalizes itself. Manus has internalized this completely. Manus runs on Claude with its custom tools built for orchestration and grounding. Their agentic environment enables the agents to browse, write code, manipulate files, and execute multi-step workflows without human in the loop. They also beat OpenAI on GAIA. An interesting thing here is that they didn't build a foundation model. They built the most compatible environment for models to reason and act within. I'm coining a new term here: Situated Agency. Situated Agency is an idea that agentic capabilities are not intrinsic to the model alone, but they emerge from the coupling of a model with tools, memory, and execution environment. Manus is perhaps the first company to productize Situated Agency at scale. And now Meta owns it. Actually, this changes everything. Meta spent a lot of time struggling to build SOTA models. Llama 4 was a disappointment. Behemoth was delayed because it couldn't compete with other frontier models. They built the Superintelligence team. Acquired Scale AI. All attempts were made to close the gaps. And now the execution layer. Manus has achieved SOTA agentic performance without training a single model. They engineered the environments and let Claude handle the inference-time compute. Meta might be positioning to become an agentic infrastructure company, not a foundation model company. Meta has - > Billions of users generating real-world task data and feedback loops daily > Rayban glasses and Quest headsets as interfaces for agents > WhatsApp, Messenger, Instagram as mediums for task delegation > Zuckerberg also mentioned that he is pushing for personal superintelligence on all wearables None of this requires Meta to have the SOTA model on MMLU. It requires Meta to have the best execution environment for models to act on behalf of users. The Avocado rumours become interesting here. Avocado is Meta's tbd closed model, reportedly being developed under @alexandr_wang. If Manus's agentic systems are genuinely model-agnostic, which their architecture suggests, then nothing blocks Meta from swapping Claude for Avocado. Manus already runs Claude and fine-tuned Qwen interchangeably, routing different subtasks to different models based on capabilities. The architecture abstracts the model layer behind a smartly engineered tool-calling interface. This gives Meta a production-tested agentic environment with $125M ARR that they can gradually integrate. They inherit the execution layer, the context engineering IP, the sandboxed compute infrastructure, the customer feedback loops, then port it to Avocado when the model is ready. Things could get hot if Meta fully commits to this thesis. OpenAI is building vertically. Foundation models, custom chips, agent frameworks, consumer applications. Google is building vertically. TPUs, Gemini, search, workspace integration. Both are betting that owning the foundation model layer is essential to capturing value. Meta could be betting the opposite. If Situated Agency is correct, then the best strategy would be to build the best orchestration infrastructure. Let others race to improve the SOTA models, and swap in whatever model scores highest on your agent benchmarks at any given moment. This is how Android beat iOS in market share. Google didn't build the best hardware. They built the best platform layer for hardware makers to build on, then captured the market. Meta making the same bet on agentic AI fits with Zuckerberg's playbook. Manus may be the first sign that suggests Meta is thinking this way about AI agents. Congrats to Meta and the complete teams at Manus AI!

@evahteev & I built a Leaderboard for Cursor users based on how many tokens they burned in 2025 Link in comments

@evahteev & I built a Leaderboard for Cursor users based on how many tokens they burned in 2025 Link in comments

Guru Network Bot in your @telegram Daily crypto news digests, mindshare reports to stay ahead - and your chats and groups as knowledge base - to stop scrolling through hundreds of messages in each one. Try it now: t.me/guru_network_bot

This Wednesday we’re going live with an AMA together with @useredio We’ll unpack how Atlas helps communities automate @Telegram, run quests & leaderboards, and turn chats into a smart knowledge base.​​ youtube.com/watch?v=tuP6l-q7…

Agents that natively self-orchestrate, managing their own context, tools, and sub-agents, are the next big unlock in LLM performance. Right now, a skilled engineer building an optimized harness, with thoughtful data flow, separation of concerns, sub-agent management, etc., can make dramatic improvements over baseline for specific tasks. If a model could do this itself, that’d be a major step forward. You give it an objective and a set of tools, and it figures out the optimal way to orchestrate itself to do the task. For example, I’m building a very primitive AI scientist that I’ll open-source soon. Most of the work isn’t in the prompt, it’s in the harness… what the orchestrator sees, what sub‑agents see, what gets shared between them and when, where we summarize vs. pass raw data, and which tools each agent controls. Doing this allows me to dramatically improve what the model can do on its own. If a model can effectively design its own harness for a given problem, it’d be a huge step forward. My bet: self-orchestrating models… ones that manage their own context, tools, and sub-agents, will move the frontier almost as much as the jump from chatbot → reasoning did. Maybe more.