The layers

July 27, 2026 · View on GitHub

Fundamentals, page 1. A dated, evidence-graded map of agent engineering's vocabulary — prompt, context, harness, loop, graph — and a rubric for grading the next term before you adopt it. Standalone: you don't need compass to use the rubric. For the idea compass is built on, see the thesis; for how it's built, the five moves.

The treadmill

In 2026 the agent-engineering community named a new discipline roughly every two months.

  prompt ─────► context ─────► harness ─────► loop ─────► graph
   2023        Jun 2025      Feb–Mar 2026   Jun 2026    18 Jul 2026

That cadence is not four independent discoveries. It's the shape of content marketing riding a real trend. And the trend is real — but so is the churn, and the two are easy to confuse, because they arrive in the same blog posts.

The cost of confusing them is concrete. A practice you adopt because it was named confidently, rather than because it was measured, becomes config you can't delete: nobody remembers why it's there, so nobody dares remove it. Config nobody can justify is the agent-engineering equivalent of a commented-out block that has survived four refactors.

So before the fundamentals teach you a layer, this page teaches you how to grade one.

The map

Layer	Named	By	Governs	Evidence
Prompt	~2023	—	The wording of one request	Absorbed. Not falsified — scoped down from "the job" to one tunable component inside a larger system
Context	Jun 2025	Lütke coined; Karpathy amplified	Which tokens the model sees	Strong. Anthropic's engineering post, Chroma's context-rot finding (all 18 frontier models degrade with input length), Gartner formalized it
Harness	Feb–Mar 2026	Hashimoto named; Osmani popularized	What the system executes, permits, and repairs	Strong. Harness-only changes moved a coding agent from ~rank 30 to top 5 on Terminal-Bench 2.0 with no model change; Databricks measured 36% → 53% from harness alone
Loop	Jun 2026	Osmani named; Willison and Huntley practised it unnamed a year earlier	How many passes, judged by what, stopping when	Moderate. One arXiv treatment, wide practitioner agreement — but the term's most-quoted soundbite has no traceable source (below)
Graph	18 Jul 2026	a joke tweet	Contested — 3–4 incompatible meanings	Weak. See below

Two of these have benchmark numbers attached. One has a joke.

The two receipts

Nothing on this page matters more than these, because they are what a grading habit is for.

The fabricated benchmark. Within four days of the graph-engineering meme, a statistic was circulating — a "Microsoft, Stanford and Anthropic" study reporting 18% accuracy gains and 85% cost cuts. It traces back to an unrelated paper about industrial engineering diagrams. A second fabricated study ("$3.1M Stanford and Anthropic") was caught the same week; the writer who checked it reported simply: "It does not exist." Neither was malicious. Both were plausible, and plausible was enough.

The soundbite with no source. Loop engineering's most-quoted line is attributed to the creator of Claude Code: I don't prompt Claude anymore — my job is to write loops. It is quoted everywhere. It could not be traced to a primary source. The transcript most often cited as its origin has him saying something closer to the opposite — that he is "purely bottlenecked on how fast I can prompt" — and describing loops as producing about a third of his code, with the caveat that "it doesn't totally click yet."

The underlying practice is still sound. But a field that cannot distinguish a real benchmark from an invented one in under four days, or a real quote from a manufactured one at all, is a field where grading your sources is an engineering control, not an academic courtesy.

Grading a claim before you adopt it

Four grades. Apply them to any practice — including every one compass ships.

Grade	Means	Test
E1 — Measured	A number, from a named corpus, reproducible by you	Can you run it and get the number?
E2 — Reported	A named source with a primary link and a date; not independently reproduced	Did you follow the link, or trust a summary of it?
E3 — Argued	A coherent mechanism with no measurement	Would you notice if it were wrong?
E4 — Asserted	Confident, popular, unsourced	Who first said it, and when?

The test column is the load-bearing part. Most E4 claims survive because nobody asks the question in the right-hand column — and every one of them reads exactly like an E2 until you do.

Two rules follow, and they're what separate this from a reading exercise:

Never let an E3 or E4 claim write config. Argument and assertion are fine for deciding what to try. They are not sufficient for a rule that will sit in every future context window, unexamined, for a year.
A claim's grade is a property of your verification, not of its author. Anthropic publishing something doesn't make it E1 for you. Following the link and running the corpus does.

What this says about graph engineering

Applying the rubric to the newest rung, honestly:

It began as satire — "Are we still talking loops or did we shift to graphs yet?" — mocking this exact treadmill. Earnest essays followed within hours, a fabricated benchmark within days. The term now carries at least four incompatible meanings: multi-agent orchestration topology, graph-structured knowledge and memory (which genuinely predates the meme by years and has real research behind it), execution traces, and "graphs of loops." And the technical objection is the one nobody has answered: a loop is already a directed cyclic graph. Even LangChain, whose product is literally named for the idea, published the deflation rather than the victory lap — a loop is "a simple version" of a graph, not its rival.

Grade: E4 for the discipline, E1–E2 for the older knowledge-graph work now being conflated with it. Those are not the same claim and should not inherit each other's credibility.

One thing underneath it is real and modest: a node in an orchestration graph can now be an entire autonomous agent run to completion, not a single model call. That genuinely changes failure isolation and per-node cost accounting. It did not need a new discipline name, and compass will document it under orchestration when it earns an E1.

What is actually load-bearing

Strip the naming and one distinction survives every source, uncontested:

What the model sees (context) is a different engineering problem from what the system around it enforces, measures, and repairs (everything after context).

That's the split worth teaching. The finer subdivisions — harness vs. loop vs. graph — are still being argued over by the people who coined them, and they disagree about the ordering: one puts loop above harness, another nests context inside harness, and Anthropic's own flagship context-engineering post never uses the word "harness" at all.

The rest of the fundamentals series teaches the durable parts and dates the rest.

Honest limits

This rubric grades evidence, not importance. An E4 idea can be correct — the practice usually arrives before the measurement, and Willison and Huntley were running loops a year before anyone named them. Grading tells you what you're entitled to assert, and what you should attach a date and a source to. It does not tell you what to try.

The grades here are also a snapshot, and a self-interested one: compass's own thesis sits on this map. loop-engineering.md was written independently of the June 2026 essays and arrives at the same structure, which makes compass a participant in this argument, not a neutral referee. Read the sources, not the summary — every claim above links to one.

The takeaway

The treadmill will produce another name before this page is a quarter old. That's fine. Grade it, date it, and make it earn an E1 before it earns a line of your config. A discipline you can't cite is a discipline you can't delete.