Shtander

Full-time reader of API invoices. AI, technology, and the numbers underneath.

[Shtander @RomanShtander]

this is free gold sitting in the docsAnthropic published their own agentic workflow patterns on their engineering blog last year. no launch tweet, no announcement, quietly in the documentationthe company published the exact patterns that make their model production-ready. that part is on the recordthe quote doing the rounds — "graph engineering makes agents 1000x more reliable" — has no source anywhere, so i'm not passing it off as factwhat's public and checkable: Anthropic's own "Building effective agents" cookbook
prompt chaining — decompose task into sequential steps
routing — classify input, hand off to specialized paths
parallelization — split work across simultaneous agents
orchestrator-workers — one agent breaks down and delegates
evaluator-optimizer — one generates, one critiques, loopfive patterns, free, sitting in the docs while the timeline waits for GPT-5learn them this weekend and by monday you're building production agent systems, not demos. that's the skill getting handed AI architecture work this year

[Shtander @RomanShtander]

@eng_khairallah1 this is stunning

[Shtander @RomanShtander]

@KobeissiLetter Tech and oil are taking it all — the consumer sector is left behind. 📉

[Shtander @RomanShtander]

@aistasiia the result is fire, but better not show anyone the path to it

[Shtander @RomanShtander]

@yurshevv anyone could buy, sitting for 50 years — no one

[Shtander @RomanShtander]

@thegreatest_sv 14 days to a real result - clear plan👏

[Shtander @RomanShtander]

@hanakoxbt Great breakdown. Memory without curation just bloats context for no benefit

[Shtander @RomanShtander]

@Nekt_0 That's why every model upgrade needs testing on real edge cases, not just benchmarks

[Shtander @RomanShtander]

@zerohedge Not surprising. The job market reacts to AI faster than the curricula do

[Shtander @RomanShtander]

@0xCodila separating "thinking" from "execution" is a smart idea

[Shtander @RomanShtander]

@0xKiyoro $20 → $2,353 in a day is impressive, but that usually means it can vanish just as fast 🙂

[Shtander @RomanShtander]

@VitalikButerin Personal air filtration and ventilation. Quiet protection for health.

[Shtander @RomanShtander]

@TheAhmadOsman 100% the right model for the task > the most powerful model for everything

[Shtander @RomanShtander]

The number that matters most from the new note on production AI agents:

44 points.

That's the benchmark swing on the same model, just from changing the harness. Zero model upgrade.

Not prompt tuning. Not RAG. This is about the infrastructure the model runs inside: guides, sensors, memory, permissions, observability. All 6 layers.

The identity this note makes explicit: Agent = Model + Harness.

Your "waiting for GPT-5" tweets made sense in 2024. In 2026, the bigger gain comes from rewriting the harness around the GPT-4 or Claude 3.5 you already have.

Three eras of AI engineering, per the authors: Prompt (2023-24) → Context (2025) → Harness (2026). Each one subsumes the last.

95% of agent startups die in pre-production. Not because the model is weak. Because nobody built the harness.

The next edge isn't in the model. It's in the runtime around it.

[Shtander @RomanShtander]

@qwinsi0x Weather markets + a DeepMind model — an unexpected combo. Cool.

[Shtander @RomanShtander]

@N01ennn Rare to see someone actually read the primary sources. Thanks for the breakdown 👏

[Shtander @RomanShtander]

@KobeissiLetter November's going to be a hot month for the AI market

[Shtander @RomanShtander]

@KobeissiLetter No cost at all — that's exactly the part worth double-checking 🤔

[Shtander @RomanShtander]

@undefinedKi $500 saved is serious. But is it really all automatic?

[Shtander @RomanShtander]

@Argona0x Sounds incredible