Agentic Technologies GmbH
Age of
Intelligence.
Cognition is becoming infrastructure.
We build the companies, products, and research that run on it.
We build with agents.
For you, for us, in public.
A small Zurich team with a large agent workforce. Three lines of work, one operating system underneath.
Build
Agent systems for your company.
We find where your team loses hours, then design, deploy, and run the agent layer: research, content, monitoring, operations. Wired into your data and your controls.
- Map
- Pilot
- Run
Products
Tools we run on ourselves first.
Everything we build for clients runs in our own operation before it runs in theirs. What survives daily use becomes a product.
- Use
- Harden
- Ship
Research
Briefs for operators and capital.
Mechanisms, numbers, second-order effects. What changed in AI and what it does to cost, speed, and headcount. No hype feed.
- Read
- Test
- Publish
Thesis
Capacity used to scale with headcount.
Now it scales with design.
For a century, what a company could do was bounded by whom it could hire and coordinate. Language models cut that link. Research, drafting, monitoring, and analysis now cost compute, not calendars.
What stays scarce is architecture. Which loops run on their own, what context they share, where each decision lands and who owns it. That is the work we do, for clients and on ourselves first.
Compute
Capacity you rent, not hire
Context
Memory that compounds
Design
The part that doesn't commoditise
An operating layer,
not a chat window.
Memory, execution, feedback, and orchestration, designed as one system. Six properties everything we build is held to.
Institutional Memory
Context compounds.
Decisions, threads, and data feed a living model. Less re-briefing. More continuity across people and agents.
Execution Layer
Work moves without waiting.
Background jobs, monitors, drafts, follow-ups. The queue clears overnight, not at the next stand-up.
Thin Overhead
Signal over ceremony.
Agents coordinate at machine speed. Humans stay in the loop through tight reviews - not meetings for their own sake.
Feedback Loops
Rated work gets better.
Approve, edit, reject - then feed the signal back. The stack improves from real use, not slide decks.
Model Freedom
Best brain for the job.
Frontier APIs or open weights. Swap the model, keep memory and process. No single-vendor religion.
Multi-Agent
Specialists, one desk.
Research, ops, code, comms - separate agents with shared state. Orchestrated, not a single chat blob.
What we're watching.
Substance briefs - mechanisms, numbers, second-order effects. Built to be useful to operators and capital, not a hype feed.
What a token per second feels like.
Four speeds, named after things you already use. Faster rows write more, so you can actually see them.
The model emits one token at a time. At this rate you either read every word as it lands or you watch a wall arrive. Throughput is not intelligence - it is the difference between a tool you wait on and a tool you work with.
The model emits one token at a time. At this rate you either read every word as it lands or you watch a wall arrive. Throughput is not intelligence - it is the difference between a tool you wait on and a tool you work with. A chat window at reading pace feels like a person thinking next to you. You can interrupt, reread, change the ask. Past about fifty tokens a second the sentences complete before you finish the first one. You stop reading and start skimming for the answer, the number, the decision.
The model emits one token at a time. At this rate you either read every word as it lands or you watch a wall arrive. Throughput is not intelligence - it is the difference between a tool you wait on and a tool you work with. A chat window at reading pace feels like a person thinking next to you. You can interrupt, reread, change the ask. Past about fifty tokens a second the sentences complete before you finish the first one. You stop reading and start skimming for the answer, the number, the decision. ChatGPT in a normal reply sits in that middle-fast band: whole sentences land, then whole paragraphs. Gemini Flash is a step past that - cheap, quick, a wall of useful text if you asked for a wall, and a wall of noise if you did not. The rate does not make the model smarter. It only decides whether you wait on the tool or work with it. Quantize it, batch it, cache it: the tokens still come out one after another. Local on a laptop CPU you wait. Local on a Mac you keep up. On a frontier API you scan. On a flash API you catch the shape and move on. That is what a token per second feels like.
The model emits one token at a time. At this rate you either read every word as it lands or you watch a wall arrive. Throughput is not intelligence - it is the difference between a tool you wait on and a tool you work with. A chat window at reading pace feels like a person thinking next to you. You can interrupt, reread, change the ask. Past about fifty tokens a second the sentences complete before you finish the first one. You stop reading and start skimming for the answer, the number, the decision. ChatGPT in a normal reply sits in that middle-fast band: whole sentences land, then whole paragraphs. Gemini Flash is a step past that - cheap, quick, a wall of useful text if you asked for a wall, and a wall of noise if you did not. The rate does not make the model smarter. It only decides whether you wait on the tool or work with it. Quantize it, batch it, cache it: the tokens still come out one after another. Local on a laptop CPU you wait. Local on a Mac you keep up. On a frontier API you scan. On a flash API you catch the shape and move on. That is what a token per second feels like. None of this is batched server throughput. One request, one stream, characters landing in order. If the row feels slow, that is the point. If the row feels like a document appearing, that is also the point. The model emits one token at a time. At this rate you either read every word as it lands or you watch a wall arrive. Throughput is not intelligence - it is the difference between a tool you wait on and a tool you work with.
Names are approximate - typical single-stream replies, not a lab bench. 1 token ≈ 4 characters. Not batched server throughput.
Get the briefs first.
Substance over hype. New insights land in your inbox as we publish them - mechanisms, numbers, second-order effects.