https://cairocananea.com.br/en/tags/ai-agents/ Recent content in Ai Agents on Cairo Cananea Hugo -- 0.152.2 en Thu, 16 Jul 2026 00:00:00 +0000 https://cairocananea.com.br/en/projects/bloom-multi-agent/ Thu, 16 Jul 2026 00:00:00 +0000 https://cairocananea.com.br/en/projects/bloom-multi-agent/ Adapting Stanford's Motivational Interviewing coach to self-directed learning: an LLM-driven Coordinator-Specialist multi-agent system with a ReAct calendar tool loop, MCP-based Google Calendar sync, and SQLite long-term memory — built for agency, privacy, and supportive recovery instead of rigid streaks. https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-4-memoria-modelos-validacao/ Wed, 15 Jul 2026 00:00:00 +0000 https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-4-memoria-modelos-validacao/ Fourth and final article in the series on building Bloom, a multi-agent coaching agent, in one week. Covers long-term memory without vector search, the non-linear behavior of small models when swapping the underlying LLM, and the evaluation harness (LLM-as-a-judge, layered testing, per-call telemetry) used to validate every change. https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-3-prompt-antes-arquitetura/ Tue, 14 Jul 2026 00:00:00 +0000 https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-3-prompt-antes-arquitetura/ Third article in the series on building Bloom, a multi-agent coaching agent, in one week. Shows how to diagnose the right layer before overhauling architecture, prompt techniques that reduce conversational rigidity (information-goal states, few-shot examples, soft limits), the importance of honestly modeling the 'unknown', fallback path design, and why behavioral safety needs to be hard-coded, not left to the prompt alone. https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-2-loop-agir-sem-sair-do-controle/ Mon, 13 Jul 2026 00:00:00 +0000 https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-2-loop-agir-sem-sair-do-controle/ Second article in the series on building Bloom, a multi-agent coaching agent, in one week. Shows how to separate read tools from write tools, why every agentic loop needs an iteration ceiling, how to isolate control signals from the response text, the effect of latency in multi-call loops, and how minimal permission scope and MCP sandboxing protect a real Google Calendar integration. https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-1-quem-decide-codigo-ou-modelo/ Sun, 12 Jul 2026 00:00:00 +0000 https://cairocananea.com.br/en/posts/2026/07/bloom-capstone-1-quem-decide-codigo-ou-modelo/ First article in a series on building Bloom-for-Learning, a multi-agent study coaching agent, as the capstone for the AI Agents: Intensive Vibe Coding course (Google and Kaggle). Covers the hub-and-spoke architecture with stateless specialists, LLM-based routing with a delegation ceiling, the most expensive bug in the project (a deterministic state override by the model), the hybrid deterministic guard over agentic routing, and the difference between tool-facing and human-emotional logic.