Most coaching apps prescribe rigid schedules and punish missed sessions. Bloom-for-Learning adapts Stanford’s Motivational Interviewing coaching model into a multi-agent LLM system with a ReAct calendar tool loop, MCP calendar sync, and local-first memory — co-creating study plans instead of dictating them.
The last article in the series on building Bloom in a week: how to give an agent long-term memory without a vector database, what actually changes when you switch to smaller local models, and how to validate changes in a probabilistic system with evidence, not gut feeling.
Third article in the series on building Bloom in a week: why most of the perceived rigidity in a multi-agent agent comes from the prompt layer, not the architecture, and how to hard-code behavioral safety invariants.
Second article in the series on building Bloom in a week: how to give an agent real freedom to act in the world — writing real events into Google Calendar via MCP — without giving up control over when and how it acts.
First article in a series on building Bloom-for-Learning in one week: why the most important decision in any agentic system is deliberately defining what belongs to deterministic code and what belongs to the model’s judgment.