Capital de Tokens
Ideas, analysis, and independent articles from the channel and its episodes.
Blog RSSA metacognition framework for LLMs proposes measuring correctness, conflict, experience, importance, and emotional load to choose between a fast path and multi-agent deliberation. The important idea is not giving AI consciousness, but making reasoning observable, measurable, and controllable.
A real migration from classic Application Insights URL Ping Tests to Standard Tests using Codex with GPT-5.6 Terra, Azure CLI, Azure MCP, and ARM, preserving alerts, secrets, end-to-end validation, costs, and rollback.
OpenAI plans to remove its models from Cursor on November 12, 2026 following the SpaceX acquisition. Beyond the corporate conflict, the episode exposes a key technical risk: depending on one model provider inside an agentic harness.
Technical debt makes software more expensive to change; operational debt makes it more expensive to keep running. We explain the difference, how both accumulate, how they reinforce each other, and why observability migrations such as URL Ping Tests → Standard Tests are a good example of both.
A viral 2023 tutorial showed how to put ChatGPT inside Jupyter by reusing the browser session. In 2026 the pattern is still valuable, but the implementation has changed: Jupyter AI, the official API, local models with Ollama, and executable notebooks turn the LLM into a true analysis agent.
A video by Gustavo Entrala shows email, calendar, presentations, spreadsheets, and automations controlled from ChatGPT. The key is not a spectacular demo: it is the architectural shift toward Work, apps, Cloud Browser, and Scheduled Tasks, where natural language becomes an interface for executing real work.
The primary interface may stop being Jira, Slack, or Stripe and become your agent instead. We analyze the agent-first paradigm shift: less time inside apps, more use of their capabilities through MCP, WebMCP, and APIs.
Microsoft took Intelligent Terminal from an experimental fork with integrated chat to a multi-agent platform: local models, OpenCode, per-tab agents, execution inside WSL, ACP sessions, and a roadmap pointing toward persistent terminals and dynamic commands. We review what changed between July and August 2026 and what is still under development.
LEAP 71 tested rocket engines and an aerospike generated by Noyron, a computational engineering model that encodes physics and manufacturing rules. We compare the video's claims with technical sources and explain what really changes when software starts designing complex hardware.
Physical AI is not simply running a model on a robot. We analyze the edge-first architecture required by the physical world and break down microsoft/physical-ai-toolchain: ROS 2, LeRobot, Isaac Lab, Cosmos, OSMO, Azure ML, SiL/HiL evaluation, ONNX/TensorRT, GitOps, and agents.
Repositories no longer contain only code: they also contain instructions agents read. We analyze how repository poisoning works, why the model alone is not enough, and which defenses modern harnesses use: instruction hierarchy, sandboxing, firewalls, secret isolation, least privilege, and human review.
WebMCP proposes that web applications publish structured capabilities for AI agents directly from the browser. We analyze its architecture, how it differs from MCP, what advantages it offers over computer use, and its security and adoption risks.
Capital de Tokens
