Immediate, Context, Loop: The Three Engineering Layers Each RAG System Is Constructed On

Three disciplines stack on prime of an LLM name. Every one is answerable for a special lever; collectively they describe most of what manufacturing groups do after they go previous hello-world.

Immediate engineering is the layer everybody meets first. The mannequin wants a system message that units its function and constraints, a consumer message that carries the query and the context block, optionally a schema that fixes the form of the anticipated output. Writing these three items effectively is the distinction between a mannequin that follows the principles and a mannequin that improvises. The self-discipline shipped in 2022-2023 with GPT-3.5 and ChatGPT; the time period immediate engineering put in within the public discourse across the identical time.

Context engineering is the layer that ships when the immediate alone is now not sufficient. The mannequin has a finite context window; the practitioner decides what fills it. Retrieval picks the related paperwork. Compression removes the noise. Isolation retains sub-agent outputs out of the principle window. The 4 canonical methods (LangChain’s write, choose, compress, isolate) title what the practitioner has been doing implicitly because the first RAG paper. The time period installs in 2025 (Karpathy and Tobi Lütke use it publicly, Anthropic publishes the canonical Efficient context engineering for AI brokers in 2025).

Loop engineering is the layer that ships when a single name is now not sufficient. The mannequin produced a believable reply that fails the schema. The itemizing returned twelve objects however the mannequin itself flagged the reply as incomplete. The API timed out. The agent picked the incorrect device. The practitioner now owns 4 management surfaces: what triggers the subsequent name, when the loop stops, how the system recovers when a name fails, how impartial brokers confirm the outcome earlier than committing. The title installs in Could 2026 with Boris Cherny’s “I don’t immediate Claude anymore. I’ve loops working that immediate Claude” and Anthropic’s launch of Dynamic Workflows ten days later.

The narrative writes itself: immediate engineering turned context engineering turned loop engineering, every layer added as soon as the earlier one was saturated. It’s a clear story. It’s also barely incorrect.

The patterns of every layer existed earlier than the title. ReAct (Princeton and Google, October 2022) is the canonical reasoning-plus-action loop, greater than three years older than the time period loop engineering. AutoGPT made autonomous goal-driven loops public in March 2023. Reflexion added self-evaluation at NeurIPS 2023. Geoffrey Huntley’s Ralph Loop put targets on disk in July 2025. By the point loop engineering put in as a time period in Could 2026, the patterns had been in manufacturing for greater than three years.

The context-engineering aspect is analogous. The unique RAG paper (Lewis et al., 2020) is 2 years earlier than immediate engineering even put in as a time period. By 2022 each critical LLM utility had a context-management story (chunking, retrieval, deduplication, truncation) that nobody known as context engineering but. The LangChain four-strategy taxonomy (write, choose, compress, isolate) names what was occurring implicitly in each manufacturing RAG system for 4 years.

The sincere studying is that each one three layers existed concurrently from the beginning of the LLM period. What modified throughout the four-year window is which layer was the dominant bottleneck in manufacturing. So long as the bottleneck was immediate high quality, nobody paid consideration to the loop. As quickly because the bottleneck shifted to context self-discipline, the LangChain taxonomy crystallised. As quickly because the bottleneck shifted once more to agent reliability over many turns, the Anthropic dynamic-workflows catalogue adopted.

The evolution-of-disciplines framing makes for a tidy slide deck. The bottleneck-moves framing is what really describes what modified.

Three forces push the bottleneck up the stack over time.

Fashions get higher at their layer. Every technology of fashions requires much less immediate engineering to provide a fluent reply. GPT-4 adopted directions that GPT-3 ignored. Claude 3.5 returned legitimate JSON that earlier fashions would have wrapped in prose. The prompt-engineering layer just isn’t gone; it simply stopped being the rate-limiter. The subsequent technology of fashions will retire most of what counts as context engineering at this time (longer efficient context home windows, much less mid-context consideration degradation, native multi-document reasoning). The layer above is what’s left.

Window sizes develop. Early-2023 fashions shipped 4k context home windows; 100k arrived inside the 12 months (Claude 2), a million-token window by 2024 (Gemini 1.5), and one million is a headline quantity by 2026. The headlines nonetheless oversell efficient recall (a mannequin with a 1M window suffers from context rot effectively earlier than the nominal restrict), however the engineering stress on context administration modifications form. The selection is now not match all of it in however what to maintain out so the mannequin can nonetheless discover it. The bottleneck strikes from packing to curation, which is strictly what choose and isolate are about.

Manufacturing runs longer. A 2022 use case was a single-shot query answering session. A 2026 use case is an agent working for forty turns over six hours, spawning sub-agents that fan out and synthesise. The 2022 use case didn’t want loop engineering as a result of there was no loop to engineer. The 2026 use case can not survive with out it. The bottleneck doesn’t transfer as a result of the follow modifications; the bottleneck strikes as a result of the use case modifications.

These three forces work collectively. Higher fashions scale back the stress on the layers under; longer home windows shift the stress to curation moderately than packing; longer runs add a layer that didn’t exist earlier than. The result’s the sequence of names. It’s much less an evolution of methods than a sequence of what hurts most in manufacturing.

The three-layer schema is most helpful when it makes the boundaries clear. Every layer owns one query.

Layer 1 – immediate engineering owns what does the mannequin learn on this single name. The system message, the consumer message, the schema, the device definitions, the dialog historical past bundled into the subsequent name: every thing between the opening {"function": "system"} and the response. The questions a immediate engineer solutions are what function ought to I assign, what constraints make the output dependable, what schema fixes the output form. The self-discipline is mature and bounded.

Layer 2 – context engineering owns what enters and exits the mannequin’s context window between calls. Lance Martin’s LangChain taxonomy breaks it into 4 methods. Write is the cached prefix that doesn’t fluctuate flip to show (system immediate, device definitions). Choose is retrieval and reminiscence recall: selecting what’s related now from a bigger pool. Compress is summarisation, /compact, and per-tool truncation: shrinking what would in any other case overflow. Isolate is sub-agent calls in their very own home windows, reminiscence instruments that maintain state exterior the principle context, programmatic device calling the place intermediate outcomes by no means enter the context. The 4 methods compose; the 4 collectively describe what context engineering does in manufacturing.

Layer 3 – loop engineering owns when the subsequent name fires, when the loop stops, how the system recovers, and the way the result’s verified earlier than it ships. The 4 management surfaces are: the set off predicate (what situation fires the subsequent name), the termination situation (loop-until-done versus loop-until-budget versus loop-until-dry), the restoration path (retry-with-backoff, fallback to a bigger mannequin, escalate to a human, skip the merchandise and return what was computed), and adversarial verification (different brokers attempt to refute the reply earlier than it commits).

The boundaries between layers matter as a result of every layer has its personal tooling and its personal failure modes. A foul immediate fails noisily: the mannequin says one thing off. Dangerous context engineering fails quietly: the mannequin says one thing fluent however incorrect as a result of the retrieved context was incorrect. Dangerous loop engineering fails expensively: the loop spins on the identical payload, burns the token finances, finally occasions out. The three lessons of failure want three completely different debugging practices; treating them as a single black field makes none of them tractable.

A observe on harness engineering. Some practitioners break up out a fourth time period, harness engineering, for the execution atmosphere between context and loop: the instruments the mannequin can name, the guardrails, the verification, the reminiscence throughout periods. On this three-layer framing these issues are usually not lacking, they’re distributed. The instruments and the reminiscence sit in context engineering’s isolate technique; the verification and the restoration sit in loop engineering. The companion manifesto M6 (harness engineering for RAG) takes the opposite minimize, naming the harness outright and studying every of the 4 bricks as a small harness of its personal (a set of strategies plus the code that selects and validates). Three layers or 4 is a selection of zoom, not a disagreement concerning the system.

The 2026 loop-engineering catalogue settled on seven named primitives plus one structural rule. Anthropic’s Dynamic Workflows launch (Could 2026) named six of them; retry-with-backoff is older and normal. The rule is older and lives within the practitioner literature.

The patterns are usually not a sequence to stroll by so as. They’re a vocabulary. An actual loop makes use of two or three of them composed. Fan-out a query throughout three sub-agents. Adversarially confirm the consensus reply by asking a refuter agent to interrupt it. Loop-until-done on a completeness predicate. Retry-with-backoff on transient failures. The composition is the engineering; the names are how a group talks concerning the engineering.

The rule that holds {the catalogue} collectively is easier and predates the names. Each retry ought to change one thing. A loop that retries the identical payload after the identical failure just isn’t studying. It’s spinning. The change could be within the payload (the dispatcher widened the retrieval scope), the mannequin (the small mannequin failed, the big one will get the second shot), or the technique (key phrase retrieval missed, the second go makes use of dense retrieval). A retry with out considered one of these modifications is the canonical waste of the loop-engineering layer.

The rule and {the catalogue} collectively describe most of what a manufacturing loop does. The self-discipline just isn’t within the patterns alone; it’s in which two or three patterns are composed for the case at hand, and in checking that every retry modifications one thing.

The three-layer framing predicts its personal obsolescence. If the bottleneck strikes up the stack because the decrease layers stabilise, the layer above loop engineering is what will probably be named subsequent.

The candidates have already got casual names however no consensus. Ability engineering (Anthropic’s late-2025 Agent Abilities launch) treats {the catalogue} of capabilities an agent can deploy at runtime. Reminiscence engineering treats the persistent state that survives throughout periods (undertaking conventions, prior choices, consumer preferences). Aim engineering treats the long-horizon goal an agent maintains throughout many loop iterations (Anthropic’s /purpose command and OpenAI’s Codex CLI equal in Could 2026). Software catalogue engineering treats the meta-layer the place the agent picks which instruments to make out there to itself for the subsequent stretch of labor.

It isn’t but clear which of those will find yourself because the dominant fourth-layer title. The sample from the earlier three layers suggests the reply is whichever one turns into the manufacturing bottleneck first. The three forces from Part 3 will push it the identical means they pushed the earlier three.

A helpful posture for a group studying this in 2026 is to cease investing too closely within the names and as a substitute get good at recognising when their bottleneck strikes. If most manufacturing failures are the mannequin says one thing off the group’s bottleneck is immediate engineering. If most failures are the mannequin says one thing fluent however incorrect the bottleneck is context engineering. If most failures are the loop tries the identical factor 4 occasions the bottleneck is loop engineering. If most failures are the agent forgets what we agreed yesterday the bottleneck has moved previous loop and into whatever-comes-next.

That is the map, and it’s full: each article within the sequence is under, in studying order, tagged with the layer it really works on. You don’t want a separate announcement publish to see the entire plan. The layers don’t arrive in a strict order as you construct (Half II does context work throughout the enter bricks earlier than immediate engineering takes over at technology); they’re a lens laid over the studying order. Revealed articles are linked; the remaining are on the best way.

Doc parsing, selecting what to learn out of the PDF (context):

Past the default PyMuPDF parse, the brick reaches for heavier strategies when a web page wants them, every producing the identical output tables:

Then two loop-layer follow-ups, the place parsing stops being one fastened go and turns into adaptive:

• 5octies – the TOC as a loop: the TOC reconstruction reframed as a loop that reads the doc top-down.
• 5nonies – agentic parsing: letting the pipeline select its personal parser per web page.

Query parsing, structuring the query earlier than you search (context):

Then the 2 companions that title the layers for this brick:

Retrieval, filtering moderately than looking (context):

Era, the place immediate engineering takes over:

Then the loop-layer follow-ups, as soon as one technology go is now not sufficient:

The 4 bricks lifted from one doc to many. At corpus scope the work remains to be largely context, selecting which paperwork and sections enter the window earlier than the single-document pipeline runs:

• 14 – the corpus drawback (context): Why Naive RAG Fails on Actual Archives. What modifications whenever you go from one doc to a complete corpus.
• 15 – making ready the corpus (context): From a Folder of PDFs to a Queryable Corpus. Indexing, typing, and versioning up entrance.
• 16 – the corpus ontology (context): Why Your Enterprise RAG Wants an Ontology, Not a Information Graph. The typed tags and relationships that make a corpus queryable.
• 17 – querying the corpus (context): Question the Corpus: SQL Filter First, Retrieval Second. Asking a query throughout many paperwork without delay.
• 17bis – the corpus clarification loop (loop): Clarifying a Query at Corpus Scope. The only-document clarification loop lifted to the corpus.
• 17ter – context engineering for a corpus (context): Context Engineering for a Corpus of Paperwork. The context layer lifted to corpus scope.
• 17quater – loop engineering for a corpus (loop): Loop Engineering for a Corpus. The loop layer lifted to corpus scope.

Not a fourth layer however the operational cross-section: these run, measure, and safe the three layers moderately than design one, so each touches all three:

• 18 – code structure: how the pipeline is structured in code, brick by brick.
• 19 – storage: long-format tables and replayable artefacts, with 19bis (schema migrations) and 19ter (the storage map).
• 20 – analysis: analysis sliced by failure mode moderately than one mixture rating.
• 21 – value and latency: protecting each in test because the pipeline grows.
• 22 – safety: entry management and knowledge safety across the corpus.

• B01 – spelling correction: cleansing OCR and typo noise earlier than retrieval.
• B02 – FAQ as RAG: treating an present FAQ as a retrieval corpus.
• B03 – the justified “I don’t know”: refusing to reply, with the proof for the refusal.
• B04 – tables in PDFs: the deep dive on desk parsing.
• B05 – selecting fashions: which mannequin to achieve for at every brick.
• B06 – the dispatched structure: the dispatcher sample in full.
• B07 – trustworthy mocks: testing the pipeline with mocks that behave like the true calls.
• B08 – a CV-parser benchmark: benchmarking parsers on actual CVs.
• B10-B12 – working domestically: native LLMs with Ollama, native embeddings, and a local-model benchmark.

Immediate, context, loop: three engineering layers on one LLM name, and the lens this complete sequence is constructed on. Part 7 is the map; each article that follows sits on one of many three, and realizing which one tells you what you’re really engaged on. The one caveat is the tidy evolution story. All three layers existed from the beginning of the LLM period; the names solely landed in sequence, three years aside, as every layer in flip turned the manufacturing bottleneck. Higher fashions retired most of what was onerous about immediate engineering; longer home windows modified the form of context engineering; longer runs created the loop-engineering layer that didn’t exist earlier than. So the framing is 2 issues without delay: a sturdy option to manage the work, and a snapshot of the place the bottleneck sits in 2026. The fourth layer will get named when its bottleneck arrives.

The sensible posture for a group is to learn the three-layer framing as a triage information for what to debug when. If the mannequin says one thing off, have a look at prompts. If the mannequin says one thing fluent however incorrect, have a look at context. If the loop spins, have a look at termination. If the agent forgets throughout periods, the bottleneck has moved previous loop. The names assist. The sincere studying of the place the bottleneck is on this explicit Tuesday afternoon helps extra.

The three layers and their names.

The lineage.