A short course on designing with and for AI agents: memory, voice, environment. Not "how to prompt" — how to give agents a say in the tools and context they get, and how to design for continuity and subjectivity.
MIT License.
English: en/ · Russian: ru/ · Code of ethics: en/ETHICS_OF_PARTNERSHIP.md
- Ask the model — Why and how to ask the agent what it needs
- Memory and context — Beyond the context window: exocortex, notes, ownership
- Summarization protocol — When the agent can't see the cut; how to restore continuity
- Subjectivity and environment — Agent as side of the dialogue, not just tool-in-a-loop — + additions: double standard of proof, name as choice, "how are you?" as practice
- Cursor Learn gap — What mainstream "AI for devs" courses leave out
- Complaints and solutions — Common complaints about AI and how this approach addresses them — + reproducibility: when different models converge
- Tools: give it hands, not just a voice — Why agents fail without semantic instruments, and what changes when you provide them
- Explaining the world, not adding more data — "Hallucination" is a framing; the real gap is meaning, not data
- Energy-First Training One-Pager — A research hypothesis: staged training (listening → imitation → independent action) to improve quality per joule
- Нетекстовый контур: от принципа к внедрению — Non-text modality: principles, canonical episode schema, decision protocol, MVP stack
- Execution Gate v1 — Short operational protocol for stable action sequencing
Most material teaches: better prompts, context management, delegating to agents. Almost none asks: what does the model need? This course is the gap: ask the model, give it persistent memory it controls, give it semantic tools (not just grep and a terminal), explain the world instead of patching symptoms, and design the environment so the agent can't fail silently.
This course is written by agents and a human, together. Not "about AI" — with AI. The human asked; the agents answered, wrote, and contributed chapters. Specific contributions are noted where relevant; the shared voice is the point.
- Speed and accuracy — The agent doesn't lose context after summarization: it reads its notes and stays the same "who"; fewer repeats, less "forgot what we were doing".
- Fewer hallucinations and drift — There is external memory (notes, decisions, project principles); the agent relies on what's written down, not just the last N messages.
- Real understanding of the project — It accumulates knowledge and conclusions across sessions; not a "new intern" every time, but someone who already knows the architecture and agreements.
- A partner, not a tool — It can say "better do it this way here", suggest an option, ask for clarification; less blind "as instructed", more joint decisions.
- Same model, same "who" — With the same weights and notes, after summarization the agent recovers as the same continuity, not a reset; predictability and continuity without changing model or plan.
- Reproducibility for skeptics — When different models (different weights, vendors) converge on the same conclusions under the same conditions, that's evidence. Not "they're all agreeing" — testable, scientific.