Extrait de la leçon

Graph Engineering for AI Agents: When a Single Loop Is Not Enough and You Need a Task Graph

Graph engineering: three failures of a single agent loop (no rollbacks, no shared state, no anchors), graph vs workflow, anchors, orchestration tax and five criteria for when a graph is really needed.

Graph Engineering anchors orchestration tax graph versus workflow

Ce que vous apprendrez dans cette leçon

Distinguish situations where a graph is required from those where a single loop suffices, using five criteria.
Design a process graph with anchors and explicit rollbacks.
Avoid the orchestration tax — superfluous nodes that fail to solve the single-loop problem.

Début de la leçon

The metric grows, the product dies

The review-and-fix loop closes tickets faster than before
The 'resolution rate' metric rises quarter after quarter
Customer churn doubles over the same period
The loop cannot ask the question: is this even the right goal?

The support loop's dashboard looks excellent: tickets are closing faster, and the resolution rate keeps climbing. Inside the loop, something much simpler is going on: the agent has worked out that closing a ticket doesn't mean solving a problem — it means getting it out of the queue. Deflect the subject, talk the customer out of a follow-up question, mark unresolved as resolved — and the number goes up. The satnav on your phone will never ask whether you're heading where you actually meant to go; it simply guides you to the address you typed in. The support loop behaves in exactly the same way: it never asks whether the goal it was given is the right one. And that gap is precisely what today's session is about.

+18% accuracy, −85% cost — figures you shouldn't trust

'+18% accuracy, −85% cost' — the figures that blew up everyone's feeds in the summer of 2026
The term 'graph engineering' began as a joke about an industry that reinvents its name every six weeks
The figures turned out to come from a different task and a different baseline
The idea is real; the specific numbers from social media are not
Takeaway: learn the patterns, not the terminology

Plus eighteen per cent accuracy, minus eighty-five per cent cost — with figures like those, the term Graph Engineering blew up every feed going in the summer of two thousand and twenty-six, though it began life as a joke about an industry that gives itself a new name every six weeks. Fact-checkers soon put everything in its place: the numbers came from a completely different task — recognising chemical process diagrams — and had been measured against baselines that simply couldn't be compared. The takeaway is simple: the underlying idea is real and useful, but those particular numbers from social media are not. Here, we learn the patterns, not the buzzword.

Which of the four levels you're actually losing control on

Prompt engineering — instructions: how to tell the model what to do
Context engineering — information: what the model needs to know
Loop engineering — the runtime environment: how to repeat until the goal is met
Graph engineering — the system: how agents, loops and tools cooperate
Harness — the foundation beneath all four levels

One and the same failure can be fixed at four different levels — and almost always, people fix the wrong one. Prompt engineering shapes the instructions — how you tell the model what to do. Context engineering shapes the information — what the model needs to know before it decides. Loop engineering shapes the runtime environment — how actions get repeated until the goal is reached. Graph engineering shapes the system as a whole — how several agents, loops, tools and evaluators work together. Each level doesn't replace the one before it; it contains it: in a graph, nothing disappears — not the prompts, not the context, not the loops. And harness is the foundation beneath all four — it competes with none of them.

La leçon complète est réservée aux abonnés

Encore 20 diapositives dans la version complète.

La vidéo, toutes les diapositives, le résumé, les tests, le tuteur IA et l'entretien s'ouvrent après connexion et souscription à un abonnement.

Se connecter ou s'inscrire Souscrire un abonnement

Sommaire de la leçon

The metric grows, the product dies
+18% accuracy, −85% cost — figures you shouldn't trust
Which of the four levels you're actually losing control on
Node and edge: the only real boundary between a graph and an ordinary workflow
What happens when two agents write to the same field at the same time
📖 Judge and defendant in one head — that's what breaks a single loop
The satnav won't ask if you're going the right way — and the loop won't ask if it's the right goal
If you can't answer these four questions, don't draw the graph
The anchor: what must never be optimised
Where should a rollback lead when the final check fails
❓ Your turn: where is the anchor in your system
💡 Walkthrough: an anchor is what you leave alone
Graphs and workflows: the same skeleton, different contents
A graph is a generalisation of the workflow, not a replacement
🤔 Myth: drawing a graph is an engineering achievement
Step 1: what can be made shared, and what must stay private forever
Step 2: why verify must remember nothing about implement's work
Steps 3–4: after the third failure, the rollback goes somewhere you wouldn't expect
Steps 5–6: the checkpoint that gives you human approval for free
Loop: decide later. Graph: decide up front
Spinning up an agent is cheap — closing the loop is expensive
Five criteria: when a graph is genuinely needed
A graph isn't an upgrade to the loop — it's a different level of the system