"Graph engineering" clearly explained (and how it can get you more out of claude code/codex) Watch:

Greg does it again!

FREE CLEAR EXPLANATION OF GRAPH ENGINEERING HELPS YOU GET MORE OUT OF CLAUDE OR CODEX AI

Graph engineering gets practical when it maps retry loops, tool outputs, and failure states. Everything Claude Code can do in 2026 (and how to actually use it)

prompt engineering, context engineering, now graph engineering. the naming layer is the only part of the AI stack that ships on schedule.

This is the bit most people skip. I locked the daily agent jobs into a fixed graph: watch, thread-check, draft, bank. Without the edges named, they invent shortcuts, and you get confident nonsense. The fix was not a bigger model. It was forcing the handoffs before the prompt.

This graph engineering workflow is great. One challenge is keeping track of the various agents and the documents they create, so you can edit them. Here's a demo of one way to do that.

Graph structure makes context retrieval way cleaner

I watched the episode. Then I went to Claude and created a Graph for creating podcast episodes. I went through a test run, and before it was ready to write the script, it gave me this message: "This run has earned some MEMORY.md entries: the graph caught a fabricated citation, a wrong legal claim, and an inflated broker count before they reached your microphone." That was pretty nice. I'll record the script soon and run the edit pass. I'll post the episode this Friday. Pretty good stuff, thanks again.

The thing this actually fixes is retrieval. Agents lose the thread across files because flat context gives no stable address to return to. What decides whether it holds up is maintenance. A stale graph is worse than none, since the agent trusts it.

The real unlock is using the graph to represent task dependencies so claude stops re solving context it already resolved. Have you tried anchoring shared context nodes vs regenerating them per task?

I broke my product planning into connected subproblems instead of one massive prompt and the output quality jumped immediately. Each solved piece feeds context into the next one.

Back your winners. Cut your losers. Greg Isenberg and Graph engineering are two things Skenai adopted early. Thank you Greg. Keep the great content flowing!

I tried this on a messy codebase last week. Instead of telling Claude what to build, I told it how I think about the problem. Completely different output.

The shift feels similar to prompt engineering a couple of years ago. The advantage is moving from what you ask to what context the model can reliably traverse.

ran the workflow from this talk in claude code to see what it prints before it spawns anything. it shows the plan first - the four phases, the subagent prompts, the file list - and waits on a human. then it names the part the talk leaves out, in yellow: parallel subagents burn tokens against your usage limit. that warning is the tool's own, not mine.

Interesting video, please check out this new tool called give me your thoughts.

The useful graph is the one that lets you rerun only the failed node instead of paying for the whole conversation again.

graph engineering sounds like a phrase someone made up to explain what they're already doing

This is one of those concepts I think more people will be talking about in a few months

This maps hard onto deal work. A rent roll, a T-12, a loan doc and an LOI are not four files, they are one graph with the same asset in the middle. Dump them in a folder and the model answers each one correctly and still misses that the tenant on page 3 is the reason the coverage ratio breaks in year two.

Don't worry. You don't have to do anything new. This happens behind the scene.

Interesting point. We've seen the same issue when building browser automation. Turning live browser sessions into reusable APIs reduced a lot of repeated work.

does the graph mainly reduce repeated context setup in Claude Code/Codex?

Agent graphs move work. Knowledge graphs connect information. I’m exploring the missing handoff with Boron Context: an ontology-first, local-first layer that gives each run sourced context and preserves verified state for the next. MIT + pre-alpha. Blunt feedback welcome.

yeah and once you get the concept, the shift in how you actually use claude code is where you place the decision gates, not just the task nodes. i kept stuffing routing logic into task prompts and by hop 3 everything had drifted with no clean trace back.

The only thing I'm seeing anywhere is ego engineering

The term "graph engineering" sounds like something I'd accidentally promise a client in a meeting, then spend the weekend learning.

The part that clicked while building one: most of the graph should not be a model. 22 step kinds in my runner, only 4 are billed as model calls. The rest is loops, gates, compute, subrecipes. Graph engineering is mostly deciding which nodes do not get to think.

Graph Engineering is tech term for Content Harnessing ;-)

Graph engineering turns claude code from scattered guesses into a connected map

nobody with actual results calls it graph engineering, they just describe what they did with their prompts

Well done @gregisenberg! Great content, perfect depth, length, and detail. Thank you for sharing.

Genuine question: when the graph upstream context changes after you already built work on a downstream node, do you rerun everything below it or just patch that node? That part always trips me up in Claude Code and Codex sessions.

Curious if this beats prompt hacks.