Open vs Permissioned AI
LangChain launched Deep Agents Deploy as an open‑source, model‑agnostic deployment option while Anthropic restricted access to its Mythos model to large tech firms on cybersecurity grounds — a clear split between open and permissioned AI approaches. For crypto projects building agentic identity, coordination and payment rails, that divergence strengthens the case for open‑agent infrastructure where users retain memory and control. (blockchain.news) (infobae.com)
Two artificial intelligence companies just made opposite bets on the same week of April 2026. LangChain put its new agent deployment system into beta as open source on April 9, while Anthropic kept its new Mythos model behind a small-gate rollout for selected big firms after warning it could help cyberattacks. (langchain.com) (cnbc.com) An artificial intelligence agent is software that does jobs in steps instead of answering one prompt at a time. LangChain says Deep Agents is built for “long-running tasks” like planning, managing context, and splitting work across subagents instead of acting like a single chatbot reply. (langchain.com) Deploying an agent means turning that software into a service that can run all day, remember prior work, and connect to tools. LangChain’s documentation says Deep Agents Deploy exposes more than 30 endpoints and supports memory, human approval, and open protocols like Model Context Protocol and Agent Protocol. (docs.langchain.com) LangChain’s pitch is not just “open source” in the GitHub sense. Its April 9 post says the system is model-agnostic, works with any model that supports tool calling, and can be self-hosted so the agent’s memory stays on the developer’s own infrastructure. (langchain.com) (docs.langchain.com) Anthropic made the opposite case with Mythos. CNBC reported on April 7 that Claude Mythos Preview was designed to find weaknesses and security flaws in software, and Anthropic limited access to companies including Microsoft, Amazon, Apple, CrowdStrike, Palo Alto Networks, and Cisco instead of releasing it broadly. (cnbc.com) Anthropic wrapped that rollout inside Project Glasswing, a locked-down testing program for defensive cybersecurity work. Forbes reported on April 9 that Anthropic described Mythos as powerful enough to spot serious software vulnerabilities, which is useful for defenders and dangerous if attackers get the same tool. (forbes.com) That split is really a fight over where the “brain” of an agent lives. LangChain says memory can stay with the user or company that runs the system, while Anthropic’s restricted rollout shows a model maker deciding who is allowed to use the strongest capabilities and under what conditions. (docs.langchain.com) (infobae.com) That difference matters most for agents that need a long memory. LangChain’s Deep Agents overview says these systems can persist memory across conversations and threads, use durable stores, and hand tasks to specialized subagents, which makes the stored history part of the product rather than a disposable chat log. (docs.langchain.com) For crypto projects, that stored history starts to look like identity. If an agent is handling wallet permissions, recurring payments, or group coordination, the valuable asset is not just the language model but the memory layer that records preferences, approvals, and prior actions across time; that is an inference from how LangChain describes persistent memory and self-hosting. (docs.langchain.com 1) (docs.langchain.com 2) So the story this week is not “open beats closed” or “closed beats open.” It is that April 2026 produced two clear templates at once: one where agent infrastructure is portable and user-controlled, and one where frontier capability is permissioned and distributed like a sensitive security product. (langchain.com) (cnbc.com)
