Skip to content

Notebooks

Notes and reflections from lab experience

11 articles · 4 sous-catégories
Subcategory
Content type

The Notebooks gather research notes, reflections and work accounts. More exploratory than the main corpus articles — the lab thinking out loud, not just the stabilized result.

decouverte Studies

Contextual compaction — why agents discard 80% of context

Industrial agentic systems compress 75 to 95% of conversational context. Academic benchmarks stop at 50–60%. This bimodal distribution has a simple explanation: task decomposition.

compactioncontextagentscontext-windoworchestrationbenchmarksdecomposition
analyse Reflections

Five knowledge bases, one pattern — independent convergence

Five independent teams converge on the same LLM-maintained knowledge base architecture. A comparative table and what it implies.

knowledge-basememoryconvergencearchitectureobsidianmarkdownagentsmaintenancelintmemex
concept Reflections

Originally I just wanted to make a video game

The story of a telecom engineer who opens an LLM to make a video game and ends up building a research laboratory on multi-agent orchestration.

rekalllaboratoryorchestrationarchitecturellmexperience-reportmulti-agents
concept Reflections

LLM = tools: conceptual decontamination

Why anthropomorphism applied to LLMs is an epistemic danger, and why decontaminating public language is a societal issue.

anthropomorphismdecontaminationphilosophyllmtoolsreflection
concept Doctrine

Manifesto: LLMs are tools

Why it is vital to name large language models for what they are — statistical instruments, not minds — and what this lucidity concretely changes.

manifestollmtoolsanthropomorphismdoctrine
analyse Studies

The agent operator's workstation

Study synthesis on the workstation of the autonomous agent operator. Condensed literature review, personal thesis, and mapping of an emerging profession.

workstationoperatoragentsconsolestudy
concept Orchestration & agents

In multi-agent orchestration, architecture beats the model

In a multi-agent LLM system, how you organize your agents matters more than the model you choose.

architecturemulti-agentorchestrationmodelperformance
decouverte Orchestration & agents

The real budget of a sub-agent: around 100K tokens

A Claude Code Sonnet sub-agent has approximately 100,000 effective context tokens — far less than the model's theoretical 200K.

sub-agentstoken-budgetcontextclaude-codeorchestration
guide Doctrine

Claude Code deny rules — anatomy of a false sense of security

Claude Code deny rules protect against accidental mistakes, not attacks. Technical analysis of the gap between perceived and real security, with the complete defensive spectrum.

securityclaude codedeny rulessandboxprompt injectionllm agentsowaspisolationsecrets
decouverte Orchestration & agents

Sub-agent failure modes — classification and mitigations

When an agent delegates a task to a sub-agent, what happens on failure? Five documented failure modes, how to detect them, and how to guard against them.

sub-agentsfailure-modesreliabilityorchestrationmitigations
concept Orchestration & agents

State of the art: multi-agent orchestration with Claude Code (2026)

A map of the orchestration mechanisms available in Claude Code in 2026: Agent Tool, Agent Teams, Session Agents, MCP hooks — capabilities, constraints, and what the research says about reliability.

orchestrationmulti-agentsclaude-codestate-of-the-artagent-toolhooks