Made By Agents

Agentic coding consultancy. We audit your processes and build the AI system in your business. Videos and notes on coding agents, models and tools.
By @tobiaswup

[Made By Agents @MadeByAgents]

Start with one boring workflow that happens 100 times a week. Measure it. Then scale.

[Made By Agents @MadeByAgents]

They're not pulling back on AI. They're pulling back on AI without a plan.

What we usually find when a company says "AI didn't work for us":

Nobody defined what a correct output looks like.
Nobody measured the process before, so nobody can prove the result after.

[philip lewis @Phil_Lewis_]

Employers worldwide are pulling back on AI as they navigate a flood of errors and low-quality work that sap employee time, raise privacy and security risks, and drive up costs https://t.co/CeCNYfhKSb

[Made By Agents @MadeByAgents]

Our rule when we build AI systems: "If a step is predictable, it's code. Not a prompt."

PDF to text → code converts
Text to line items → code parses
"Which product does the customer actually mean?" → that's the LLM's job

Fewer LLM calls = cheaper, faster and you can audit it

[Made By Agents @MadeByAgents]

7. Who owns it when something breaks?

Missing 3 of these? Then it's a demo, not a system.

This is the checklist we run in every deep AI audit.

[Made By Agents @MadeByAgents]

Every week someone tells us they built an AI agent over the weekend and it's ready for customers.

Before an agent touches a real quote, invoice or customer email, we check 7 things:

[Made By Agents @MadeByAgents]

1. Does it pass a test set of real past cases?
2. Does a human approve anything above a risk threshold?
3. Is every input, output and tool call logged?
4. What happens when it's unsure?
5. Is there a cost limit per run?
6. Can it only see the data it actually needs?

[Made By Agents @MadeByAgents]

https://t.co/h0JJCgmlyR

[Made By Agents @MadeByAgents]

All major open-source System One decision models are now listed on our website together with relevant benchmarks 🏆

Happy exploring Jev alternatives 🤓

Link ⬇️

[Made By Agents @MadeByAgents]

In depth video about how to use Jev with coding agents is coming online the next days!

[Tobias Wupperfeld @tobiaswup]

Jev vs locally running Laya-MLX and Kev-4B

I built my own snake benchmark. Jev @typesafeai is running via API.
The other models are small alternatives running on very little RAM on my Macbook!

Jev seems to deliver the best quality no doubt! After running it for a while 0 deaths.
Laya died pretty early on and then ran in a loop without catching any food.
Kev died twice in the same time but runs much faster at ~175ms latency while Jev needs ~395ms.
Laya is crazy fast at ~33ms but I guess you need to train it to increase output quality.

[Tobias Wupperfeld @tobiaswup]

This new model runs on consumer laptops or Mac Mini.

The Q4_K_M variant by @UnslothAI needs only 7.12GB of VRAM to load the weights.

[Tobias Wupperfeld @tobiaswup]

I think a lot of companies will start shifting towards local AI inference just because of the cost.

AI subscriptions are heavily subsidized at the moment. Once that money is gone, many companies will realize it's cheaper to hire humans than to run agents.

The next months will be interesting. It's like with Bitcoin: the one that has cheap energy and compute has a clear advantage.

I'm working on several local AI app projects right now @A1Laboratory and I see an exciting future for that.

[Tobias Wupperfeld @tobiaswup]

LLMs can't fix broken inputs. They just hallucinate over them.

That's why the OCR layer is so important when building anything with documents.

You don't want your agents to approve wrong invoices or misinterpret contract terms.

Bad text in → confident wrong decision out.

I built a fully open-source document intelligence pipeline that solves this permanently.

↓ here is the breakdown

[Tobias Wupperfeld @tobiaswup]

Your knowledge base is blind.

It can't see charts, diagrams, screenshots, or designs.

Every image in your company is invisible to search.

@GoogleDeepMind just changed that.

Their new Gemini Embedding 2 puts text, images, PDFs, video, and audio into ONE unified vector space.

So I built a full-stack multimodal RAG on top of it.

Upload a photo of a chip → it surfaces the exact spec sheets, reports, and videos that match.

No workarounds. High speed. High retrieval quality.

I open-sourced the whole thing 👇

[Tobias Wupperfeld @tobiaswup]

Every AI model leaderboard is lying to you.

Not on purpose. They just only measure one thing: raw benchmark performance.

Which means a 70B model that barely edges out a 7B on MMLU gets ranked higher, even if it needs 5x the VRAM and costs 10x more to run.

That's not a ranking. That's a spec sheet.

So I built the MBA Open Score, a 0–100 composite that ranks every model on what actually matters:

→ Benchmark performance (40%)
→ Community adoption (25%)
→ Performance per GB of VRAM (20%)
→ Versatility (15%)

The efficiency dimension alone will surprise you.

A 7B model that scores 72 on benchmarks while running on 6GB of VRAM can outrank a 70B scoring 74 on 80GB.

Benchmarks are theory. VRAM is reality.

[Tobias Wupperfeld @tobiaswup]

Let me introduce the Agent Framework Wizard.

Helps you pick the right framework for building out your team of AI agents.

Compare up to 3 frameworks side-by-side.

Capabilities, strengths and trade-offs at a glance.

Free to use 👇

[Tobias Wupperfeld @tobiaswup]

I built a free tool that lets you compare the costs of running your AI agents.

Buying a GPU vs renting a GPU vs paying for AI APIs

What should you actually do?

This tool projects the total cost of each option over your chosen time horizon and picks the cheapest.

[Tobias Wupperfeld @tobiaswup]

I seeded a range of new AI models into my website.
Synced all benchmarks.

Very interesting to see what's trending right now.

Gemini 4 is absolutely ripping it!
Qwen3 family also gaining popularity.

You can run these on 11GB of VRAM comfortably or rent a GPU for a few cents/hour.

[Tobias Wupperfeld @tobiaswup]

Renting GPUs shouldn't be hard.

I built a live price index across @runpod and @vast_ai normalized per-GPU. Refreshed hourly. No affiliate weighting.

Cheapest H100 right now: $1.17/GPU-hr.
Run open-source models at full speed.

What providers should I add next?

[A1 Lab @A1Laboratory]

Have you ever wondered if your device can run local AI? 🧐

The AI Hardware Calculator tells you exactly which open source models your device can handle before you download anything.

@tobiaswup ran it on a MacBook and got instant results on speed, quality, and fit. 🔥👇