Catalog

What a language model actually does, what the context window costs you, and which knobs change its output.

How a general model is born in pre-training, and the two ways you specialise it afterwards.

Giving a model knowledge it was never trained on, using embeddings and a vector store.

How a model reaches outside itself, the mechanism underneath every agent.

Three kinds of memory, parametric, working and long-term, and where each one actually lives.

The agent loop, and what turns a single model call into multi-step work.

When one agent stops being enough, the two architectures to know, and the two problems you inherit.

Traces are the only record of what an agent actually did, and the loop that turns them into a better agent.

Writing the input that makes the output you want the most likely one, and why the job title disappeared.

Deciding what earns a place in a limited context window, and why more context makes answers worse.

Why an agent that can write and run code can do almost anything, and the eight standard ways you extend one.

The mechanical scaffolding around a model that keeps a non-deterministic agent on track, and why the same model scores differently in different harnesses.

Designing the loop that keeps an agent working until a goal is met, and what happens when nobody is left to press enter.

Planned. Its schedule of parts:

• Generative UI: the model producing the interface, not just the text inside it
• Event streaming: sending a run's steps to the front end as they happen
• Server-Sent Events, and why they fit an agent better than a single response does

How people get models to produce what they should not, and the guardrails, guard models and red-teaming tools used to stop them.

Agents that live with you rather than in a repository: lifelong memory, self-written skills, and reachable from the chat apps you already use.

Planned. Its schedule of parts:

• RAG vs. Agentic Search
• JSON vs. CLI tools
• Frontend Tools

Planned. Its schedule of parts:

• Cognee
• MemSearch
• Hindsight

Planned. Its schedule of parts:

• A2A
• Context delegation vs. subagent context delegation vs. messaging pool
• Shared context against isolated context, and what each costs

Planned. Its schedule of parts:

• Reflexion
• Tree of Thought
• LLM Council

Planned. Its schedule of parts:

• Superpowers
• SpecKit
• GSD

Planned. Its schedule of parts:

• Claude Code best practices, and the CLI tools worth reaching for
• Rewind and resume: going back to an earlier point in a run rather than starting again
• Skills that a subagent runs from the command line

Planned. Its schedule of parts:

• Harness profiles
• System prompts
• Renaming/reshaping tools for a given context

Planned. Its schedule of parts:

• THREAD
• ReAct
• CodeAct

Planned. Its schedule of parts:

• Agentic UI
• Generative UI
• Event streaming

Planned. Its schedule of parts:

• Agent servers
• LangChain Agent Protocol

Planned. Its schedule of parts:

• Self-supervised learning: how pre-training gets a training signal with no labels
• Post-training: everything done to a base model before it is released
• Reinforcement learning, and RLHF

What an agent framework actually gives you, the three layers of the LangChain stack, and when a workflow tool is the better answer.

Who runs the model for you: the gateway, the first-party vendors, and what actually differs between them once you are paying.

The software that actually runs the weights, when you should care, and what changes once you are serving more than yourself.

Getting an interface out of a model, and the one file that stops every generated screen looking like a different product.

Planned. Its schedule of parts:

• Every protocol mentioned in this series
• NLWeb
• UCP