TL;DR
The post (Charly Wargnier, LinkedIn, 2026-11-04) is a pointer, not an argument: it recommends secondbrainos.dev, a free, MIT-licensed course and handbook set by Yarchi (repo) that renders 130 Markdown pages, a 57-term glossary and 100+ tools as one force-directed graph (1,236 links, about 9-10 per page). The "other way to understand AI" is twofold. The content says an agent is the model plus five layers around it (context, loop, gate, harness, evals), each answering one question, and most failures live in a layer, not the model. The form says learn the field as a dependency graph, not a linear syllabus, because a failure in one layer masquerades as another.
The mental model: five layers, one system
The site's Five layers, one system page frames the agent as a bright new hire on day one; everything that decides success is the system built around them. It rejects treating "context engineering", "loop engineering" and so on as rival schools: they are layers of one stack, and each answers a single question.
| Layer | Question | Analogy | Failure when missing | Looks like |
|---|---|---|---|---|
| Context | What does it see? | What's on the desk | Guesses, invents, re-asks for things it was given | A dumb model |
| Loop | Who decides the next step? | Checklist vs. self-directed | Brittle script, or a wanderer burning tokens | Flakiness or cost |
| Gate | Who sorts incoming work? | Front desk sorting mail | Junk and trivia hit the most expensive path | Runaway cost |
| Harness | What can it reach? | Office, keys, reviewer | Can decide but not act, or acts on everything unchecked | Useless or dangerous |
| Evals | How do you know? | The same test every month | Regressions found by users | Nothing, until prod |
Underneath sits Google's agent anatomy (model = brain, tools = hands, orchestration loop = nervous system, deployment = body) and its five levels of agency (0 bare model to 4 self-evolving); the site says most production value today is at levels 1-2. The diagnostic heuristic worth keeping: name the layer before blaming the model. The "Gate" layer is the least standard of the five; it maps to cheap classifiers or small "System One" models returning typed decisions in front of an LLM.
Engineering claims worth keeping
A loop has four parts or no brakes. A goal with a programmatic definition of done, a checker outside the model (tests, compiler, schema; self-review doesn't count, citing Huang et al., ICLR 2024), a stop rule (pass, turn cap, or two identical attempts), and a budget in both turns and dollars. The model chooses the next step but never votes on whether it's done. Route every non-pass exit to a deterministic fallback.
Build the harness outside-in: containment, guides, sensors, permissions. Each ring must hold when every inner ring fails. Containment (VM, worktree, read-only DB user, network allowlist) needs zero model cooperation. Spend approval prompts only on the irreversible: Anthropic reports users approve 93% of Claude Code prompts, so frequent prompts train people not to read them. Hooks turn a guide into an enforced rule.
Harness components are bets that expire. Example: Anthropic added context resets for one model generation and removed them for the next. The site prescribes a quarterly re-audit.
Tools are where agents quietly fail. Fewer, bigger tools; descriptions are prompts; return meaning, not payloads; errors should teach the fix; typed tools beat raw access (the schema is containment). MCP's tool-description channel is an injection surface (tool poisoning).
Context is a budget, not a bin. Decide what is always in, on demand, or never; compact; treat the filesystem as long-term memory and the window as working memory.
The second-brain project (Karpathy's April 2026 LLM-wiki pattern) uses three rules: nothing is ingested until linked, contradictions are recorded with dates and sources rather than overwritten, and no vector DB until grep plus links stops being enough.
Critique
The taxonomy is the real contribution: it gives a shared vocabulary for root-causing agent failures, and it is grounded in primary sources (Google's agent whitepapers, Anthropic engineering posts, cited papers) with a contributing rule that every factual claim needs one. The graph UI is a good onboarding device, but it is presentation; the value is in the Markdown.
Weaknesses to keep in mind:
Single-author curation. One maintainer's synthesis, recently published (the README still has a pre-publish TODO). Treat the "New this month" lists and star counts as a snapshot checked 2026-11-01, not ground truth.
The gate layer is opinionated. It leans on "Jev", a specific small decision-model product, for a whole handbook. The principle (cheap classifier in front of the big model) is sound; the vendor emphasis is not neutral.
Thin on multi-agent and governance. Multi-agent is a level in the taxonomy but not a layer; there is little on authority, audit, or who may change agent memory, which R002 flags as an open problem.
The post adds nothing beyond the link. 84 reactions and 7 comments by midday, all short praise (one compares it to Obsidian's graph view); no pushback to learn from.
Using it
A layer-by-layer review checklist for any agent design:
Context: can you list what is always in the window, what is fetched on demand, and what is excluded?
Loop: is "done" a program that returns true or false, and is the checker outside the model?
Loop: are there a turn cap, a dollar cap, a spin detector, and a deterministic fallback?
Gate: which requests could a rule or small classifier handle before the big model sees them?
Harness: what is the blast radius if every prompt-level guide is ignored?
Harness: do approval prompts fire only on irreversible actions?
Evals: is there a fixed task set, including known failures, run on every model or prompt change?
Re-audit: which harness bets were made for an older model and may have expired?
Related entries: R002 (memory layers; the second brain's "record contradictions, never overwrite" is one answer to R002's open contradiction problem). R004 (both say config rots across model upgrades; this site turns R004's one-off wipe into a scheduled audit with evals).
Sources
LinkedIn post, Charly Wargnier, 2026-11-04. Includes a 39 s screen recording of the site.
secondbrainos.dev (not reachable from the cloud; seen only in the video).
undefined-ui/second-brain-os: README and the pages Five layers, What an agent is, The four parts of a loop, and The four rings, read 2026-11-04.