Coding agents are shaped mainly through the context placed in their limited window, so engineers should manage context instead of trying to control model internals.
2
Claude Code setups can use rules, hooks, sub-agents, skills, observers, and project-scoped files to give agents the right information at the right time.
3
A markdown-based memory wiki can preserve research, score information by use and decay, and retrieve relevant knowledge during a short coding task.
Summary
Fausto presents coding agents as systems whose behaviour depends heavily on context injection. Engineers cannot directly control the model's internal parameters, but they can control what enters its context window and when. He discusses Claude Code files, rules, hooks, sub-agents, skills, plugins, observers, and deferred tools. His practical advice is to keep context small, use project-scoped information, and start a fresh session before a context window becomes crowded. He then describes a memory system built from markdown files, an index, importance scores, decay rules, and agents that observe sessions and retrieve relevant knowledge. A demo compares an ordinary Claude Code setup with one containing a prepared wiki while both attempt to build a cellular automata visualisation under time pressure. The talk ends with a build-off in which teams extract structured information from a technical drawing into JSON. The exercise tests whether a prepared context system can help an agent solve a domain-specific task quickly.
Context injection is the part of an agent system engineers can directly change
Fausto says model internals are opaque and new models appear frequently, so engineers should focus on the context they place in the model's limited window. A raw API call is stateless, while Claude Code loads system instructions, Claude.md files, rules, skills, tools, and memory before the user starts chatting. The window may become larger and faster, but it remains finite. He frames context management as getting the best result with the minimum input, injected at the right time.
A crowded context window makes an agent less focused and more expensive
Fausto argues that context is not neutral data. Different files, concepts, and tasks compete for attention inside the same window. He recommends treating 25% context usage as a practical upper limit in Claude Code, whether the model has a 1 million token or 200,000 token window. Beyond that point, he says the system becomes more expensive, slower, and more prone to errors. His answer is to manage context through separate sessions and selective loading rather than filling the window.
Claude Code configuration should put each instruction in the mechanism that fits it
Fausto divides context into deterministic, probabilistic, and human sources. Claude.md files provide general instructions, while rules can be scoped to paths or file types. Hooks can run at lifecycle events, such as stopping a session or processing user input. He uses a separate session as an observer that rewrites dictated input and helps compensate for changes between model versions. He is sceptical of Claude's automatic memory because he finds its saving and retrieval decisions unreliable.
Sub-agents and skills let one agent delegate work without loading everything into one session
Claude Code sub-agents do not inherit the default agent's personality message, Claude files, or memory, although model and MCP settings can be inherited or changed. Fausto uses separate agents for work such as research, email drafting, invoicing, and ideation instead of sending every task to a coding agent. Skills begin as markdown instructions but can include scripts or other agents. A skill can call Gemini for native video analysis or use Codex for another task, allowing Claude Code to coordinate different systems.
Deferred tool descriptions and project scope reduce unnecessary context
A large collection of skills and MCP tools still creates a selection problem because the model needs descriptions to choose among them. Fausto recommends deferred tool loading so full tool content is injected only when needed. He also prefers project-scoped skills and MCP descriptions over putting everything at user scope. The model needs enough information to choose correctly for the current project, but loading every description everywhere wastes context. This trade-off means shorter descriptions save space while giving the model less detail for selection.
A useful memory system needs policies for what to store and retrieve
Fausto proposes a markdown wiki with raw source files, an index, summaries, concepts, entities, and procedures. Storing everything is not enough, because accumulated information needs a meaning and a retrieval policy. His experimental system lowers a concept's score when it has not been used for a period of time, while related new material can raise the scores of existing concepts. Research agents can save relevant findings for the current task while preserving other material for later retrieval.
Prepared memory can give an agent a large head start under a strict time limit
Fausto compares two similar Claude Code runs that use the same model, skills, and sub-agents. One has a wiki populated with research for a difficult cellular automata visualisation, while the other does not. The configured run can fetch relevant concepts and start building immediately. The default run must spend its short allowance researching or rely on the model's existing knowledge. Fausto uses the comparison to argue that time is an important measure of agent performance, especially when the task is deliberately too short for fresh research.
Agents should only receive tasks whose outputs can be judged
Fausto says an LLM task needs a way to determine whether an output is good or bad. That check can be deterministic or probabilistic. He doubts that writing essays or customer service are good uses in the examples he gives, while structured extraction from technical drawings is more suitable because extracted values can be checked and scored. The build-off applies this idea by asking teams to extract information from a drawing into JSON and compare the result with a known answer.
"If you are not able to say that if a model produces an output, if that output is good or bad, then you probably shouldn't give the task to an LLM."39:24
Who should watch
You are building coding-agent workflows and need practical ways to control what an agent sees without retraining a model.
Your Claude Code setup has accumulated too many files, skills, MCP tools, or memories, and you need a way to keep sessions focused.
You want to test whether a prepared knowledge base can help an agent solve a domain-specific task under a short deadline.
