KlyraA

Graph Engineering for AI agents
Nodes • Edges • Verifiers • Parallel systems

[KlyraA @klyrAaX]

@0xDepressionn Direct prompt engineering was just the training wheels. Building deterministic evaluation + context-injection harnesses is where the actual software engineering happens now.

[KlyraA @klyrAaX]

@didnobodynotice Yes, there's even more information below.

[KlyraA @klyrAaX]

$1,000,000,000 ARR in six months — that's how much revenue Boris Cherny and his team unlocked for Anthropic with a single internal tool.

Now, the creator of Claude Code himself just dropped a 28-minute masterclass breaking down their entire internal setup.

Forget basic prompt engineering tutorials. This is a complete shift in how dev teams operate:

• Zero-friction context: Using CLAUDE.md to teach the AI your architecture, linters, and code standards automatically.
• Memory shortcuts: Bypassing repetitive prompts to instantly resume complex dev sessions.
• Autonomous agent clusters: Running parallel agents across separate git worktrees to code, test, and debug simultaneously.
• Internal Anthropic patterns: The exact custom commands and workflows the creators use to save dozens of hours every week.

If you want to build at 10x speed, this short video gives you the exact blueprint to stop coding line-by-line and start managing an army of AI engineers.

Are you still using AI as a glorified chatbot, or are you orchestrating autonomous agents yet? Drop your workflow below 👇

[Anatoli Kopadze @AnatoliKopadze]

https://t.co/fyBGWQbsvn

[KlyraA @klyrAaX]

@didnobodynotice Of course, people need to clearly understand that results come from consistency, not from doing everything all at once.

[KlyraA @klyrAaX]

Google Brain co-founder Andrew Ng and Neo4j's Andreas Kollegger just showed the real breakthrough isn't more agents, it's the graph they work inside.

Andrew Ng's https://t.co/uNxvrKVdFj partnered with Neo4j to prove it, live, no cuts, building an agent from scratch on the same graph database running fraud detection and recommendation engines most of your timeline has never noticed depend on it.

The blueprint underneath it is simple: define the goal, split the work, fan out to specialized workers, verify independently, merge the results. Instead of one agent carrying the entire context and executing every step sequentially, independent tasks run in parallel while the graph controls dependencies, routing, verification, and state.

The separation is the whole point: nodes do the work, edges carry the results. If two workers don't depend on each other's output, they shouldn't be forced into the same sequential loop.

Here's where to actually watch it play out:

04:00 - he starts from a relational schema, two tables and a join, the exact structure most teams already have and never think to turn into a graph
16:00 - what actually makes an agent great, powerful through reasoning, adaptive with memory, easy to prompt, and what makes it expensive, slow remote calls, non-deterministic output, tokens that add up fast
30:00 - the agent gets its first real tool, say_hello, written and tested live, a parameterized query built specifically to resist injection instead of trusting the input
45:00 - it runs, a live conversation holding context across turns instead of forgetting the moment the message ends
Most agents don't fail because the model is weak. They fail because nobody gave them anywhere to put what they learned.

[Anatoli Kopadze @AnatoliKopadze]

https://t.co/fyBGWQbsvn

[KlyraA @klyrAaX]

@didnobodynotice But how helpful was it for you to take a break and listen? After all, this is a person who has changed and simplified life for many

[KlyraA @klyrAaX]

@didnobodynotice Adding more agents to a broken pipeline just scales the chaos exponentially. Structural control always beats brute-force compute.

[KlyraA @klyrAaX]

@0xSpecter_ Spot on. If your output loop lacks strict verifiers and isolated execution contexts, even a GPT-5 class model will eventually drift.

[KlyraA @klyrAaX]

@littledespair Yes, only clear and structured information.

[KlyraA @klyrAaX]

@HarryTandy Facts. Everyone loves drawing 10-step agentic graphs until step 3 hallucinates and breaks the main branch. Loops > Graphs every time.

[KlyraA @klyrAaX]

@0xShoopy debug UI that shows which agent handles what is the part a lot of people are still missing

[KlyraA @klyrAaX]

@sonicdr1p Exactly. The graph gives those agents the structure to manage state, run independent work in parallel, and verify results before they get merged.

[KlyraA @klyrAaX]

@elonmusk zero to usable Blender render on day one is the kind of jump that used to take months of tutorials

[KlyraA @klyrAaX]

@perfectium_xyz parallel + self-check + recovery

that’s the difference between a demo and something you can actually leave running

[KlyraA @klyrAaX]

@doc_halegaleh @GeminiApp If you think you're always too lazy to write down your opinion and back it up a little more, you're wrong, because you've gotten used to just writing a couple of words or a single sentence and that's it.

[KlyraA @klyrAaX]

@_tifa_67 I'm glad this information is helpful to you

[KlyraA @klyrAaX]

@UScap1tal This is the line worth sitting with: "no single thread drowning in context." That's the actual failure mode nobody names, teams keep upgrading the model when the real bottleneck was always what the agent was allowed to remember.

[KlyraA @klyrAaX]

@littledespair Andrew's the rare case where the hype is actually undersold. Most people teaching this stuff would've led with "revolutionary," he just quietly builds the agent and lets the schema speak for itself.