Amit Jadhav

Field Notes · Issue 003 · 07 September 2026

Everyone is still learning to prompt. The field already moved three disciplines past it.

Prompt engineering was the whole game in 2023. By 2026 it is table stakes — the floor, not the discipline. While most teams are still perfecting how they phrase a question, the people actually shipping AI have moved up a clean lineage of four named disciplines: from how you ask, to what the model sees, to what runs it, to how the whole thing is wired. And right at the top, the field has split into the argument of the year — loops versus graphs. Skip the lineage and you are optimising the one rung that stopped mattering.

Four disciplines, one rising line, and the fork everyone is arguing about — receipts, not hype, which is the whole point of Field Notes. Every claim dated and sourced to September 2026. Written for the people who have to decide, with a second layer one tap away for the people who have to build.

The spine

One rising line, four disciplines — and a fork at the top

Each discipline sits on top of the one before it; none replaced its ancestor. You still phrase a prompt inside a context you engineer, inside a harness that runs it. The argument in 2026 is only about the very top of the line — whether to reach for a loop or a graph. Tap any node to jump to it.

1 Prompt 2022–23 2 Context 2024–25 3 Harness 2026 4a Loops 4b Graphs worker level system level how you ask what it sees what runs it how it's wired

The four disciplines of modern AI engineering, 2022 → September 2026. The line rises from a single phrased request to full multi-agent systems; the dashed split at the top is the loops-vs-graphs argument. Field Notes 003.

Discipline by discipline

What each one is, why a decider should care, and what to ask your vendor

Four parts each: the plain-English version, why it matters to the business, the sourced number that proves where it stands, and the one question to put to any vendor. Switch on Builder mode for the architecture and the failure modes beneath each.

012022–23

Where you meet it

AIChatGPT
Claude
Gemini
Perplexity
PromptLayer
ChatGPTOpenAI Playground

How you ask · the ancestor

Prompt engineering

The craft of phrasing a single request so the model returns what you want. It arrived with ChatGPT in late 2022, when tiny context windows meant a few words changed everything, and for a while "prompt engineer" was a job title. It still matters — but it has been quietly demoted from a discipline to a table stake.

Why a decider should care

This is the stage almost everyone in your organisation is still standing in — typing a question, taking the first answer. It is real skill, but it is the floor of what these tools can do, not the ceiling. Treating prompting as "using AI" is how a capable team leaves most of the value untouched.

What to ask your vendor

"Show me what happens after the prompt." If the entire product is a cleverly worded box, you are buying stage one at a stage-three price.

Builder note · how it actually works

Prompting is steering a fixed distribution: system prompt, few-shot exemplars, output-format constraints, and decoding controls like temperature. The discipline decayed not because phrasing stopped mattering but because the leverage moved up a layer — the same prompt behaves completely differently depending on the context assembled around it.

The gotcha: prompt-only "fixes" don't survive a model upgrade. Anything you hard-code as a phrasing trick is liability the day the underlying model changes. Encode the intent in the harness, not the wording.

The one-line version

Prompting is the floor, not the discipline. If it's the whole product, it's the whole problem.

022024–25

Where you meet it

Claude
AIOpenAI
LILlamaIndex
PPinecone
MMCP tools
Atlan
OvalEdge
Cursor & Claude Code

What it sees · the core skill of 2026

Context engineering

Designing everything the model sees on each call — not just your question, but the state, the retrieved documents, the tool definitions and the memory. The unit of control moved from the sentence to the whole session. Andrej Karpathy crystallised it at Y Combinator in June 2025; Shopify's CEO called it "the art of providing all the context for the task to be plausibly solvable." By 2026 it was widely named the primary AI-engineering skill.

Why a decider should care

This is where accuracy is actually won or lost. A mediocre model with excellent context beats a frontier model fed noise. It reframes the buying question away from "whose model is smartest" toward "who assembles the best context around it" — and the latter is something your own team can own.

What to ask your vendor

"Where does the context come from on each call, and how is it kept fresh?" Vague answers here predict hallucinations later.

Keep > summarise The counterintuitive 2026 finding: under modern prompt caching, keeping the full history often beats summarising it on cost, latency and recall at once — because summarising rewrites the cached prefix and forces the system to pay full price to recompute what it just tried to save. OpenAI Cookbook · Anthropic Claude Code docs · production evals, 2026
Builder note · the config gotcha

Prompt caching keys on an exact prefix match. Edit or summarise anything early in the window and every token after the edit point is a cache miss — you re-pay for the whole tail. That's why compaction can raise cost and latency while it's trying to lower them.

The real failure mode: teams add summarisation to "save tokens," watch spend go up, and can't see why. The fix is an append-only history with cache-aware ordering — stable prefix first, volatile content last. Reach for compaction only when you're genuinely near the window ceiling, not by default.

The one-line version

The model is a commodity. The context you feed it is the product.

032026

Where you meet it

Claude Code
AICodex
Cursor
GitHub
LangSmith
Faros AI
Databricks
Google ADK 2.0 & MS Agent Fwk

What runs it · the third phase of maturity

Harness engineering

The scaffolding around the model that turns a clever responder into a reliable worker. HashiCorp co-founder Mitchell Hashimoto named it in February 2026 with a phrase that stuck: Agent = Model + Harness. Every time the agent makes a mistake, you engineer a permanent fix into its environment so that mistake becomes structurally impossible. Five layers are commonly cited: tool orchestration, verification loops, context and memory, guardrails, and observability.

Why a decider should care

This is the discipline that decides who ships. Frontier labs build the inner harness — native tool use, memory, safety. Your enterprise builds the outer harness: the custom configuration, routing, testing and guidelines that map a raw model onto your actual workflow. That outer harness is the real moat, and it is a hedge against lock-in — because your business logic lives there, the model underneath becomes swappable.

What to ask your vendor

"If we changed the underlying model tomorrow, how much breaks?" If the answer is "everything," you don't own a harness — you own a dependency.

The outer harness is the moat Frontier labs are actively shrinking their inner harness as models improve — which pushes more responsibility onto the outer harness the enterprise controls. The durable advantage is not the model you rent; it's the scaffolding you build around it. Hashimoto, Feb 2026 · OpenAI "Harness engineering" · martinfowler.com (Böckeler)
Builder note · the architecture & the gotcha

Concretely the outer harness is: a tool registry with typed schemas and permission scopes; deterministic verifiers (tests, linters, validators) gating each step; a memory/context assembler; guardrails (input/output filters, rate and cost caps); and full tracing. Encode it as versioned artifacts — CLAUDE.md / AGENTS.md, eval banks, config — not as tribal knowledge.

The real failure mode: treating the harness as glue code nobody owns. When it's undocumented, every model upgrade is a fire drill and every incident is unreproducible. A benchmark that exposes this: swap the model and watch your reliability numbers — if they move materially, your business logic is still trapped in the model layer.

The one-line version

You don't buy an agent. You build the harness that makes a model behave like one.

04THE FORK

Where you meet it

Linear Loops
LangGraph
Google ADK 2.0
MSMS Agent Fwk
Neo4j
LangSmith

How it's wired · the argument of 2026

Loop engineering vs graph engineering

At the top of the line the field splits. Loop engineering is the worker-level cycle — plan, act, observe, revise — governed by a verifier and a stopping condition, not by the model's own confidence. It went mainstream when Linear shipped a feature literally called Loops in July 2026. Graph engineering is the system-level move — externalising shared state and relationships so many agents can coordinate in a persistent, resumable structure.

The cleanest way to hold the difference, verbatim from the practitioner synthesis: a loop externalises revision, a chain externalises task order, a network externalises role specialisation, and a graph externalises shared state and relationships.

Why a decider should care

This is a cost-and-risk fork, not a fashion choice. A loop is cheap, legible and easy to supervise. A graph is powerful and durable but heavier to build, run and govern. Reaching for a graph before you need one is one of the fastest ways into the 40% of agentic projects that get cancelled.

What to ask your vendor

"Why is this a graph and not a loop?" A good vendor can name the shared state that forced the graph. A weak one reached for it because it demos well.

Default to the loop The rule of thumb that holds in production: if you cannot name the shared state that must survive across runs, you do not yet need a graph. Move up only when work genuinely branches, state must persist across sessions, or specialised roles need different context and permissions. Practitioner consensus + framework docs, 2026
Builder note · the architecture & the real failure mode

The tooling has converged on graphs. LangGraph is the reference stateful runtime (checkpointing, human-in-the-loop, time-travel, resumable execution). Google ADK 2.0 replaced its hierarchical executor with a graph Workflow Runtime (Python GA May 2026, Go GA June 2026). Microsoft moved AutoGen to maintenance (Oct 2025) and shipped the graph-based Microsoft Agent Framework (GA April 2026).

The real failure mode — multi-agent economics: Anthropic's own multi-agent research system beat single-agent Opus 4 by 90.2% on their internal eval — but used about 15× the tokens of a normal chat, and their published architecture had no circuit breakers or per-run caps. The 15× multiplier compounds when a sub-agent spawns more agents or a tool returns oversized output. Graphs earn their keep on genuinely parallel, dependency-light work; on tasks that need shared context they're the wrong tool.

The one-line version

Loops are cheap and legible. Graphs are powerful and heavy. Most teams reach for the graph a stage too early.

Field Notes · Issue 004 in production

One long-read a month. Nothing else in your inbox.

Each Field Note is a working page like this one — researched, sourced, interactive, and paired with a two-minute film. Subscribers get it the day it goes live.

No pitching. Unsubscribe in one click.

Reality check · Sep 2026

Four numbers that decide who actually ships

Read together, they say the same thing the four disciplines do: the model is not the bottleneck. The harness — verification, evals, governance — is.

40%+

of agentic AI projects forecast to be cancelled by end-2027 — on cost, unclear value and weak risk controls, not model quality.

Gartner, Jun 2025
23%

of organisations are scaling an agentic system anywhere in the enterprise; no more than 10% in any single function.

McKinsey State of AI, Nov 2025
90.2%

improvement of a multi-agent system over single-agent on Anthropic's internal research eval — at ~15× the tokens.

Anthropic, Jun 2025
~80

tokens at rest for an Agent Skill — the near-zero-cost, cross-vendor standard adopted by OpenAI, GitHub, Cursor and more within weeks.

Anthropic Dec 2025 · SwirlAI measurement

Notice what is not on this list: model choice. The cancellations trace to governance and cost, not capability. The multi-agent win comes with a 15× token bill that only pays off on the right problem. And the cheapest, fastest-spreading advance of the year — Agent Skills — is essentially markdown in a folder. The pattern is consistent across every credible 2026 dataset: the teams that ship are the ones who built the harness, not the ones who picked the model.

Do this next

Five moves that follow from all of the above

Each carries a threshold — the line that tells you when you've actually done it, rather than just nodded at it.

01

Stop optimising the model. Build the outer harness.

Write your tools, guardrails, verification gates and stopping conditions as versioned artifacts. Treat every production failure as a permanent harness fix, not a one-off retry.

Threshold: swap the model and your reliability numbers barely move.

02

Default to a loop; earn the graph.

A verifier-governed loop with a human review gate is correct and cheap for a single well-scoped task. Graphs are for branching work and persistent state.

Threshold: you can name the shared state that forced the graph.

03

Instrument before you scale.

Given the 40% cancellation rate and the 15× multi-agent token bill, put evals, observability and hard per-run cost caps in place before agents touch real traffic.

Threshold: no agent ships without an eval bank and a budget ceiling.

04

Adopt portable, low-cost standards now.

Agent Skills and MCP are becoming cross-vendor defaults at near-zero authoring cost. Encode repeatable workflows as Skills to cut lock-in — and vet third-party skills as a supply-chain risk.

Threshold: your top workflows live as portable Skills, not bespoke glue.

05

Cite the receipts, not the hype.

This field moves on screenshots and viral threads. Before you repeat a framing in a board deck, find the primary source and the date behind it — a named author, a dated post, a real release. If you cannot, flag it as unverified rather than passing it on.

Threshold: every claim in your deck has a source and a date.

Read this before you quote any of it

Three honest caveats

The disciplines overlap; they don't queue

This is a lineage of capability, not a calendar. You prompt inside a context inside a harness inside a loop or graph — all at once. The order tells you what depends on what, not what to retire.

Market and adoption figures are scope-dependent

"Agentic AI" has no standardised definition, so market-size and adoption numbers vary widely by analyst. Treat any single figure here as directional, and anchor to the primary sources: Gartner, McKinsey, and the labs' own posts.

Some framings are practitioner conventions

The "five harness layers," the inner/outer split, and the "loop/chain/network/graph" line are useful practitioner models, not standardised specs. And the real course is a 2025 artifact that recirculated — check the date before you cite it.

Sources & data citations — Gartner, McKinsey, Anthropic, OpenAI, Google, Microsoft, Linear, LangChain
Gartner (25 Jun 2025): over 40% of agentic AI projects will be cancelled by end-2027 on escalating costs, unclear business value or inadequate risk controls; "agent washing" flagged, ~130 of thousands of "agentic" vendors judged real · McKinsey, The State of AI in 2025: Agents, innovation, and transformation (Nov 2025): 88% of organisations use AI in at least one function; 23% scaling an agentic system somewhere in the enterprise; no more than ~10% scaling in any single function · Andrej Karpathy, "Software Is Changing (Again)," Y Combinator AI Startup School (17 Jun 2025): context engineering framed as the core skill · Mitchell Hashimoto, "My AI Adoption Journey" (Feb 2026): "Agent = Model + Harness" · OpenAI, "Harness engineering: leveraging Codex in an agent-first world" (11 Feb 2026) · Birgitta Böckeler / Thoughtworks on martinfowler.com (2026): guides-and-sensors harness taxonomy · OpenAI Cookbook, "Prompt Caching 201," and Anthropic Claude Code docs: summarisation invalidates the cached prefix; keep-everything can win on cost, latency and recall · The four agentic design patterns — Reflection, Tool Use, Planning, Multi-agent collaboration — are widely cited practitioner conventions (2024) · Knowledge-graph agent tooling: Neo4j + Google ADK patterns for multi-agent extraction and persistent state, in wide circulation through 2026 · The loops-vs-graphs debate reignited on X in July 2026 alongside Linear's Loops launch; the practical decision rule (loop for single scoped tasks, graph for persistent multi-agent state) is practitioner consensus · Linear "Loops" feature shipped 20 Jul 2026 · Anthropic Agent Skills open standard (18 Dec 2025): SKILL.md, YAML frontmatter, progressive disclosure; median ~80 tokens at rest per skill (range ~55–235), measured by SwirlAI across 17 official skills; adopted by OpenAI Codex, GitHub Copilot, Cursor, VS Code, Gemini CLI, JetBrains Junie within weeks · Google ADK 2.0: Python v2.0.0 GA 19 May 2026, Go v2.0.0 GA 30 Jun 2026 — graph-based Workflow Runtime, built-in human-in-the-loop, resumable execution · Microsoft: AutoGen to maintenance mode (Oct 2025); Microsoft Agent Framework 1.0 GA (Apr 2026) — typed, graph-based Workflow, checkpointing, pause/resume, HITL, native MCP/A2A · LangGraph: checkpointing (MemorySaver/PostgresSaver), thread-based state, HITL interrupts, time-travel debugging, resumable execution; 30k+ GitHub stars · Anthropic, "How we built our multi-agent research system" (Jun 2025): multi-agent beat single-agent Opus 4 by 90.2% on internal research eval; token usage explains ~80% of variance; ~15× tokens vs chat; no circuit breakers in the published architecture · LangChain customer story (18 Feb 2026): monday.com cut eval feedback loop from 162s to 18s per iteration (8.7×) with LangSmith on a LangGraph ReAct agent · Figures current as of 07 September 2026. Analyst market definitions vary; vendor-published tallies are directional, not audited. This page is a briefing, not legal or investment advice.

Field Notes

Get the next one before anyone else

One in-depth, interactive Field Note a month, each with a two-minute film. Written for the people who have to decide. If you'd like this walked through for your leadership team, board or campus, write to Amit directly.

No pitching. Unsubscribe in one click.