meta lessons
Topics
Living cross-source syntheses. Status advances seed → emerging → established.
Agent security
Threats and mitigations for LLM agents: prompt injection (direct / indirect), tool poisoning, data exfiltration, memory poisoning, over-broad permissions, and defense patterns (lea
Agents
Autonomous LLM agents: the agent loop (perceive - plan - act - observe), tool use, memory, planning strategies, multi-agent patterns, control flow and state, and failure modes. Con
Context engineering
Deciding which tokens reach the model , and how they are shaped: prompt authorship, context-window construction and ownership, thread/event modelling and serialisation, token budge
Evals
Production evaluation of agent pipelines: per-stage metrics (routers as classifiers, generation via pass@k and pairwise comparison), layered QA gates, golden datasets and human ali
MCP (Model Context Protocol)
The Model Context Protocol: servers, tools, resources, prompts, transport (stdio / HTTP), the client-server handshake, and how agents consume MCP capabilities. Boundary with the ne
Memory
How a system remembers across sessions : what gets written and when, how the stored thing is represented, how it is kept true as the world changes, how the human inspects and corre
Autonomous research loops
The setup an unattended improvement loop needs before it can be trusted to run for hours with nobody watching: what must be frozen and what may move; which resource is held constan
RAG (Retrieval-Augmented Generation)
Retrieval-augmented generation: chunking, embeddings, vector stores, retrieval strategies (semantic / hybrid / reranking), and grounding generations in retrieved context. Widened o
Self-improvement
The loop by which a model gets better from output it generated itself: sampling many candidates rather than one, selecting among them, and feeding the survivors back as training da
Skills
Agent skills: what a skill is and how it loads, how it is triggered, how to write one that fires when it should and not when it should not, how to evaluate one, and when to delete
Inferencing
Running models efficiently: the prefill/decode phases, the KV cache, batching (static / continuous), quantization, speculative decoding, attention optimizations (e.g. paged attenti