shopware AG
Started early, still going strong—entrepreneur at 16.
Websites
- Company Website
-
http://www.shopware.com
- Personal Website
-
https://agentic-commerce.sh/
About
Agentic Commerce | Building the infrastructure for machine customers
Shopware |…
Shopware |…
Articles by Stefan
Activity
10K followers
-
Stefan Hamann reposted thisStefan Hamann reposted thisHow to tell bots, agents and humans apart? In the agentic web, this is no longer just a bot-defense question. It shapes access to content, control over data, and eventually monetisation. At the Agentic Commerce Lab, we run a small experiment: humans, bots and agents were asked to fill a form on a honeypot site. We collected request metadata, environment clues and behavioral signals. What we found: behavioral signals, environment and metadata do tell apart humans, from agents, from bots. But not evenly. Agents that bring their own container we caught every single time. Agents running as a sidebar in your own Chrome are harder - it's literally a human's browser. In these cases, behavior was the tell: Every agent that actually filled our form gave itself away: one dispatched 401 keystrokes with 398 of them in the same millisecond, another typed at a fixed 10ms interval. Humans don't do that. But: Mostly this is a technical implementation question today, rather than a model capacity question. This can be changed and will be changed once it becomes relevant. This leaves a bigger question: who should decide which automated visitors are welcome? For an agency, and builders of crawlers and agents, access often means satisfying the policies of the infrastructure in front of the site. For a merchant or publisher, the question is different: which agents create value, which consume it, and which create risk? Web Bot Auth and Proof of Personhood point in opposite technical directions: One helps identifying automated actors. The other tries to verify humans. And again, both leave the same question: Who gets to issue the credential, and who gets to make that decision? Full Research Notes Download: https://lnkd.in/dRNEvHAQ Github: https://lnkd.in/dusj6M4j CC: Michael P. Pfeiffer Juan Fernández Stefan Hamann Sebastian Hamann Lukas Rump Bipin Babu Joachim Weber
-
Stefan Hamann reposted thisStefan Hamann reposted thisThis is our second post so please don't AI slop us yet 😄 It's Tuesday and we have some exciting news for you! Shopware's Agentic Commerce Plugin is now fully open-source. People can fork it, improve it and contribute back to the community. Its definitely worth celebrating. This release brings many improvements to the MCP protocol for UCP, making the overall setup much smoother. Find it here: https://lnkd.in/eKHgV3Et Repository here: https://lnkd.in/e2_CJXNJ #UCP #MCP #AgenticCommerce #OpenSource Björn Meyer Juan Fernández Lukas Rump Robin Schulte Sebastian Hölscher Daniel Nögel Bipin Babu Daniel Skogly Bjørn Fjellstad Corinna Bratt Nicolas Zimmel Sebastian MentingGitHub - shopware/agentic-commerce: Shopware Plugin for agentic commerce features, like Universal Commerce Protocol and native agentic discovery.GitHub - shopware/agentic-commerce: Shopware Plugin for agentic commerce features, like Universal Commerce Protocol and native agentic discovery.
-
Stefan Hamann reposted thisStefan Hamann reposted thisWhat do agents dream about? 🤖 💬 This week I went down a deep rabbit hole: agent dreaming and recursive self-improvement applied to commercial agents negotiating deals and B2B RFQs. What's agentic dreaming? “Dreaming” is just a useful human metaphor for two related capabilities: - First, memory consolidation: reviewing past interactions, discarding noise, and turning important experiences into persistent knowledge. - Second, simulation: exploring alternative outcomes and testing possible actions in a sandbox before taking them in the real world. For a negotiation agent, this means analyzing previous negotiations to understand where it conceded too early, held too long, or lost; simulating buyer responses to different combinations of price, delivery date, payment terms; identifying recurring buyer patterns; testing concession strategies against synthetic buyer agents and evaluating proposed improvements before deploying them. Now, this opens up a bigger discussion around recursive self-improvement. The interesting competitive question here may not be who builds the best agent or the best negotiation strategy, it is who builds the best autonomous learning system: a system that learns from experience, tests its own strategies, improves its own memory and tools, and evolves faster than its competitors. Imagine a supplier handling thousands of RFQs and negotiations every day. That supplier has more interaction data, more opportunities to experiment, better estimates of buyer behavior and a faster feedback loops that translates into self-improvements and a better negotiation strategy. This will create a powerful compounding advantage. In a short time, the competitive advantage may be less about access to an agent and more about access to data and the best learning loop. And that raises more difficult downstream questions! What happens to competitors with less data and fewer opportunities to learn? Do markets become more efficient, or more concentrated? Do adaptive negotiation agents create better deals, or more sophisticated price discrimination? Could agents independently converge on stable, supra-competitive outcomes? For systems like this, governance can't be an afterthought. We need transparent decision trails, human-controlled policies, clear escalation points, independent evaluation, protection against overfitting and reward hacking, monitoring, and reliable rollback mechanisms. This is one of the areas we're exploring at Shopware´s Agentic Commerce Lab. There are so many open questions, and we're still so early in this space! Interested in the topic? Reach out and say hi. 🙂 P.S. sharing some additional reading material on agent dreaming, world models, recursive self-improvement, and algorithmic competition in the comments below 👇 #AgenticCommerce #AI #RSI #AgenticProcurement #ResponsibleAI Stefan Hamann Michael P. Pfeiffer Agentic Commerce Lab by Shopware Katja Forbes Christian Magorrian
-
Stefan Hamann reposted thisStefan Hamann reposted thisAgent Sandboxing: A non-issue? When I started in the Agentic Commerce Lab earlier this year, agent permissions were in many cases still managed by "Allow" and "Allow all `ls -la` in this project". Slow and cumbersome. But letting the agent loose on your host is risky: Agents make mistakes and there are MANY stories around how agents deleted critical data. So Claude and I created a little side-project "Agent In a Box": `aib claude` or `aib zed` in any directory and I get a Debian container with relevant dev-tools pre-installed. The idea was to reuse the Linux VM every Docker daemon already runs in on MacOS, and get disposable environments without breaking dev workflows. Editors help here: VS Code or Zed both support dev containers, so this integrates smoothly. But a dev container is not a boundary itself: A forwarded SSH_AUTH_SOCK and a mounted docker socket are two isolation breakers many setups have. AIB puts an approval daemon in front of SSH agent and uses a filtering proxy in front of the docker API. Since then a lot of things have changed: Claude and Codex ship OS-level sandboxing OOTB (e.g. bubblewrap, seatbelt) with "agent supervision" if they need to leave the sandbox. And there are many dedicated agent sandboxes now, including from Docker. So a non-issue? Mostly solved, yes. Still building this taught me about dev containers, socket proxies, docker-in-docker vs proxy filtering and where isolation quietly leaks. And I think it's an interesting approach - so I wanted to share it anyway. GitHub, MacOS only: https://lnkd.in/dNve2t5Y
-
Stefan Hamann shared thisCommerce is changing faster than at any point I can remember. AI agents will search, compare, decide and increasingly transact on behalf of people. For merchants, this means a fundamental shift: their commerce infrastructure will need to serve humans and machines at the same time. I believe Agentic Commerce will become the next infrastructure layer of digital commerce. And I want to spend my time building it. That is why, after 26 years, my role at Shopware is changing. As Sebastian and I hand over the CEO responsibility to Joachim Weber, I will take on a new role as Chief Agentic Officer. This is not a goodbye. Quite the opposite. I remain on the Executive Board and a majority shareholder. But I will be able to focus much more of my energy on product, technology and the questions that made me start Shopware with Sebastian in the first place. My job as Chief Agentic Officer is now to separate hype from substance. To understand where autonomous agents can create real value for merchants. And together with my team, to turn the most relevant innovations into the Shopware platform. Joachim has my full trust to lead Shopware. Sebastian Hamann will focus on the human-led and experience side of commerce as Chief Experience Officer. And I get to focus on the machine and intelligence side. In many ways, it feels like returning to my roots. And I couldn't be more excited about what comes next. #Shopware #AgenticCommerce #AI #OpenCommerce
-
Stefan Hamann reposted thisStefan Hamann reposted thisHow to tell bots, agents and humans apart? In the agentic web, this is no longer just a bot-defense question. It shapes access to content, control over data, and eventually monetisation. At the Agentic Commerce Lab, we run a small experiment: humans, bots and agents were asked to fill a form on a honeypot site. We collected request metadata, environment clues and behavioral signals. What we found: behavioral signals, environment and metadata do tell apart humans, from agents, from bots. But not evenly. Agents that bring their own container we caught every single time. Agents running as a sidebar in your own Chrome are harder - it's literally a human's browser. In these cases, behavior was the tell: Every agent that actually filled our form gave itself away: one dispatched 401 keystrokes with 398 of them in the same millisecond, another typed at a fixed 10ms interval. Humans don't do that. But: Mostly this is a technical implementation question today, rather than a model capacity question. This can be changed and will be changed once it becomes relevant. This leaves a bigger question: who should decide which automated visitors are welcome? For an agency, and builders of crawlers and agents, access often means satisfying the policies of the infrastructure in front of the site. For a merchant or publisher, the question is different: which agents create value, which consume it, and which create risk? Web Bot Auth and Proof of Personhood point in opposite technical directions: One helps identifying automated actors. The other tries to verify humans. And again, both leave the same question: Who gets to issue the credential, and who gets to make that decision? Full Research Notes Download: https://lnkd.in/dRNEvHAQ Github: https://lnkd.in/dusj6M4j CC: Michael P. Pfeiffer Juan Fernández Stefan Hamann Sebastian Hamann Lukas Rump Bipin Babu Joachim Weber
-
Stefan Hamann reposted thisStefan Hamann reposted thisHow do you know whether an AI agent is actually authorized to make a purchase? We have added experimental AP2 support to the Agentic Commerce Plugin and published it as a prerelease. The goal was to understand how agent-led payments could work in practice. AP2 connects user intent, the checkout state, and payment authorization through verifiable mandates. Combined with the existing UCP capabilities of the plugin, this gives us a way to test a full flow from checkout creation to payment verification and order confirmation. The implementation is still early and not production-ready. We are sharing it now because working with the protocol surfaced useful questions around trust, authorization, and how commerce platforms should interact with AI agents. The video shows a simplified version of the flow. Prerelease: https://lnkd.in/efKnXtHU #AgenticCommerce #AP2 #UCP #OpenSource #Shopware Bipin Babu Bjørn Fjellstad Corinna Bratt Daniel Nögel Daniel Skogly Juan Fernández Michael P. Pfeiffer Nicolas Zimmel Robin Schulte Sebastian Hölscher Sebastian Menting Stefan Hamann
-
Stefan Hamann reposted thisStefan Hamann reposted thisIf an AI agent buys from your store, how do you know what the customer actually approved? At the Agentic Commerce Lab, we’re exploring this question by building a research preview of AP2 support for our agentic commerce plugin's UCP extension for Shopware 6. AP2 introduces verifiable mandates into the transaction flow. A Checkout Mandate binds the authorization to the exact checkout state, while a Payment Mandate provides verifiable payment authorization within defined boundaries. Together, they help protect against unauthorized changes, replay attacks, and ambiguity about what was actually approved. The video shows a complete agent-to-agent commerce transaction, including AP2 mandate verification in Shopware 6. Trust and verifiable authorization will be important as autonomous commerce moves from experimentation toward real-world adoption. #AgenticCommerce #UCP #AP2 #X402 #Shopware Bipin Babu Bjørn Fjellstad Corinna Bratt Daniel Nögel Daniel Skogly Juan Fernández Michael P. Pfeiffer Nicolas Zimmel Robin Schulte Sebastian Hölscher Sebastian Menting Stefan Hamann Shopware
-
Stefan Hamann reposted thisStefan Hamann reposted thisWe built a Shopware WebMCP integration and benchmarked it against other approaches to agentic browsing. WebMCP is an emerging standard that brings MCP into the browser. Unlike traditional MCP servers, it is tab and session bound rather than server bound. This enables entirely new use cases where users and AI agents can collaborate within the same browser session. As part of the project, we also compared WebMCP with other approaches to agentic browsing, including CUA Code and CUA Native. If you're curious about the current landscape, the video provides a practical overview of each approach. One teaser: some agents completed the same tasks nearly 5× faster while costing about 4× less than others. What I found most interesting is the tradeoff between imperative and declarative WebMCP. While the imperative approach feels like the cleanest solution from an engineering perspective, declarative WebMCP could become especially compelling if we move toward a more semantic web. If you'd like to explore the Shopware integration, it's available in our Agentic Commerce Lab's Github: https://lnkd.in/d9WpQhMN Michael P. Pfeiffer Juan Fernández Robin Schulte Lukas Rump Sebastian Hölscher Stefan Hamann Sebastian Hamann Bipin Babu Daniel Skogly Agentic Commerce Alliance
Stefan Hamann replied to a comment
1mo
Iztok Smolic 🩸 fair test! And you're pointing at the right risk: the honest worry isn't too few constants, it's whether something we haven't accounted for is quietly doing work the two numbers are getting credit for. That's exactly what the checks are built to expose — if a hidden dial is needed, the predictions should drift and the thing should break.
So I treat it as open by design: it holds until some missing factor says otherwise. Thanks for the kind words and for engaging with it seriously.
Stefan Hamann replied to a comment
1mo
Simon Neuberger Exactly — "cleaner, not gone" is the honest version. And I'd add one guardrail to the compliment: low description length is only impressive if it isn't smuggled in. A short rule that needs a long manual of hand-tuned choices isn't actually short. So the real test isn't elegance, it's whether the few bits survive contact with the data without us quietly adding more.
That's why I keep insisting on the falsifiable side: the rule either keeps regenerating the constants as we push it, or it breaks — and either outcome teaches us something. Beauty is a hint, not a proof.
Happy to pick up the philosophical thread another time. For now, agreed: back to the formal experiment, where the numbers get the last word.
Stefan Hamann replied to a comment
1mo
Dominic Klein Spot on, Dominic — that's the limitation we're most conscious of. 🙌 These are greenfield, one-shot tasks, so a big chunk of the variance really is the harness: every model gets the identical prompt and the same frozen contracts, and we score real effects rather than style. But that only answers "can it build the thing," not "can it work within ours."
The brownfield case is the more interesting one to me too — aligning with existing conventions, respecting contracts instead of reshaping them, and asking for clarification vs. silently assuming. That's exactly where we want to take the next version: drop models into an existing codebase and score adherence + restraint, not just output. Curious how you'd define "respecting an existing contract" as a measurable signal? 👀
-
Stefan Hamann liked thisStefan Hamann liked thisShopware hat gerade etwas geliefert, das viele B2B-Vertriebler jahrelang nur mit teurem Custom-Code hinbekommen haben. Mit Release 6.7.13.0 unterstützen die B2B Components jetzt Multi-Context Employee Accounts. Was das konkret bedeutet: Ein Mitarbeiter kann mehrere Rollen und Kontexte innerhalb eines Accounts abbilden. Einkäufer, Genehmiger oder Buchhaltung. Warum das für Hersteller und Großhändler relevant ist: → B2B-Kunden haben selten nur einen Ansprechpartner → Einkaufsprozesse haben oft mehrere Rollen und Freigabestufen → Genau hier wurden Shopware-Projekte in der Vergangenheit schnell komplex und teuer Mein Rat deshalb: Bevor ihr individuelle Rollenlogik entwickelt, prüft zuerst, was Shopware inzwischen standardmäßig abdeckt. Das kann Entwicklungsbudget sparen und das Budget könnt ihr an anderer Stelle sinnvoller einsetzen: zum Beispiel für bessere Inhalte, bessere Prozesse oder mehr Sichtbarkeit. Genau diesen Denkfehler sehe ich im Mittelstand immer wieder: Erst entwickeln. Dann prüfen, ob es das eigentlich schon gibt.6.7.13.0 unterstützen die B2B Components jetzt Multi-Context Employee Accounts. Was das konkret bedeutet: Ein Mitarbeiter kann mehrere Rollen und Kontexte innerhalb eines Accounts abbilden. Einkäufer, Genehmiger oder Buchhaltung. Warum das für Hersteller und Großhändler relevant ist: → B2B-Kunden haben selten nur einen Ansprechpartner → Einkaufsprozesse haben oft mehrere Rollen und Freigabestufen → Genau hier wurden Shopware-Projekte in der Vergangenheit schnell komplex und teuer Mein Rat deshalb: Bevor ihr individuelle Rollenlogik entwickelt, prüft zuerst, was Shopware inzwischen standardmäßig abdeckt. Das kann Entwicklungsbudget sparen und das Budget könnt ihr an anderer Stelle sinnvoller einsetzen: zum Beispiel für bessere Inhalte, bessere Prozesse oder mehr Sichtbarkeit. Genau diesen Denkfehler sehe ich im Mittelstand immer wieder: Erst entwickeln. Dann prüfen, ob es das eigentlich schon gibt. Wer gerade an einer komplexen B2B-Kundenstruktur in Shopware arbeitet: Meldet euch gern. Wir schauen uns den Use Case mit euch an.
-
Stefan Hamann liked thisStefan Hamann liked this576 Unternehmen in einer Stadt. So viele haben wir zuletzt in Nordhorn durch unseren Scanner gejagt. 341 davon hatten eine Website. 235 nicht. Und ehrlich: Vor ein paar Monaten hätte ich wahrscheinlich gedacht, dass die 235 ohne Website automatisch die spannendsten Leads sind. Ist aber Quatsch. Eine fehlende Website heißt noch lange nicht, dass jemand gerade eine braucht, Geld dafür ausgeben will oder überhaupt offen dafür ist. Und eine vorhandene Website heißt genauso wenig, dass alles passt. Genau das finde ich gerade spannend. Wenn man sich nicht 20 Firmen anschaut, sondern hunderte, verändert sich der Blick komplett. Plötzlich geht es nicht mehr um: „Hat Website / hat keine Website.“ Sondern eher: Warum sollte genau diese Firma gerade mit uns sprechen? Gibt es überhaupt einen echten Anlass? Ist die Website ein Problem oder einfach nur nicht schön? Passt das Unternehmen überhaupt zu uns? Bei 576 Firmen merkst du ziemlich schnell, wie viel Müll man sich in eine Leadliste ziehen kann, wenn man nur nach einfachen Kriterien filtert. Große Liste sieht erstmal geil aus. Bringt dir aber nichts, wenn die Hälfte davon eigentlich keinen Sinn macht. Das ist wahrscheinlich auch einer der Gründe, warum ich bei Vertrieb inzwischen deutlich weniger auf reine Masse gebe. Ich will nicht 1.000 Leads. Ich will wissen, welche 50 davon wirklich interessant sind. Und genau dafür bauen wir unseren Scanner gerade immer weiter aus. Nicht damit wir noch mehr Firmen finden. Davon gibt es genug. Sondern damit wir schneller merken, welche davon wir besser direkt wieder löschen sollten.
-
Stefan Hamann liked thisStefan Hamann liked thisZwei Tage Zukunft der Bildung in Ahaus 🙌 Beim gemeinsamen Workshop von Tobit AG und Manage Now haben sich unsere Kollegen Gert Mengel und Christine Bach mit einer zentralen Frage beschäftigt: 👉 Wie lernen und arbeiten wir künftig in der Schule? Von KI und Digital Leadership über Cybersecurity und Datenschutz bis hin zu modernen Lernumgebungen stand alles im Zeichen von Learnflow und der Gestaltung einer zukunftsfähigen Bildung. Danke an Ludwig Groten und Saskia Henneke für die Gastfreundschaft und die spannenden Impulse!
-
Stefan Hamann liked thisStefan Hamann liked thisThis is our second post so please don't AI slop us yet 😄 It's Tuesday and we have some exciting news for you! Shopware's Agentic Commerce Plugin is now fully open-source. People can fork it, improve it and contribute back to the community. Its definitely worth celebrating. This release brings many improvements to the MCP protocol for UCP, making the overall setup much smoother. Find it here: https://lnkd.in/eKHgV3Et Repository here: https://lnkd.in/e2_CJXNJ #UCP #MCP #AgenticCommerce #OpenSource Björn Meyer Juan Fernández Lukas Rump Robin Schulte Sebastian Hölscher Daniel Nögel Bipin Babu Daniel Skogly Bjørn Fjellstad Corinna Bratt Nicolas Zimmel Sebastian MentingGitHub - shopware/agentic-commerce: Shopware Plugin for agentic commerce features, like Universal Commerce Protocol and native agentic discovery.GitHub - shopware/agentic-commerce: Shopware Plugin for agentic commerce features, like Universal Commerce Protocol and native agentic discovery.
-
Stefan Hamann liked thisStefan Hamann liked thisWhat do agents dream about? 🤖 💬 This week I went down a deep rabbit hole: agent dreaming and recursive self-improvement applied to commercial agents negotiating deals and B2B RFQs. What's agentic dreaming? “Dreaming” is just a useful human metaphor for two related capabilities: - First, memory consolidation: reviewing past interactions, discarding noise, and turning important experiences into persistent knowledge. - Second, simulation: exploring alternative outcomes and testing possible actions in a sandbox before taking them in the real world. For a negotiation agent, this means analyzing previous negotiations to understand where it conceded too early, held too long, or lost; simulating buyer responses to different combinations of price, delivery date, payment terms; identifying recurring buyer patterns; testing concession strategies against synthetic buyer agents and evaluating proposed improvements before deploying them. Now, this opens up a bigger discussion around recursive self-improvement. The interesting competitive question here may not be who builds the best agent or the best negotiation strategy, it is who builds the best autonomous learning system: a system that learns from experience, tests its own strategies, improves its own memory and tools, and evolves faster than its competitors. Imagine a supplier handling thousands of RFQs and negotiations every day. That supplier has more interaction data, more opportunities to experiment, better estimates of buyer behavior and a faster feedback loops that translates into self-improvements and a better negotiation strategy. This will create a powerful compounding advantage. In a short time, the competitive advantage may be less about access to an agent and more about access to data and the best learning loop. And that raises more difficult downstream questions! What happens to competitors with less data and fewer opportunities to learn? Do markets become more efficient, or more concentrated? Do adaptive negotiation agents create better deals, or more sophisticated price discrimination? Could agents independently converge on stable, supra-competitive outcomes? For systems like this, governance can't be an afterthought. We need transparent decision trails, human-controlled policies, clear escalation points, independent evaluation, protection against overfitting and reward hacking, monitoring, and reliable rollback mechanisms. This is one of the areas we're exploring at Shopware´s Agentic Commerce Lab. There are so many open questions, and we're still so early in this space! Interested in the topic? Reach out and say hi. 🙂 P.S. sharing some additional reading material on agent dreaming, world models, recursive self-improvement, and algorithmic competition in the comments below 👇 #AgenticCommerce #AI #RSI #AgenticProcurement #ResponsibleAI Stefan Hamann Michael P. Pfeiffer Agentic Commerce Lab by Shopware Katja Forbes Christian Magorrian
-
Stefan Hamann liked thisStefan Hamann liked thisEin wenig klang die Zukunft der Shopware Führung in unserem Gespräch vor vier Wochen schon heraus, als Stefan Hamann sagte: "Das ist übrigens auch der Grund, warum ich und viele andere CEOs wieder zurück in die Produktentwicklung und das Engineering gehen." Shopware stellte gerade die Top-Führungsriege neu auf und macht Joachim Weber zum neuen CO-CEO.
Experience & Education
Honors & Awards
-
EY: Entrepreneur of the year 2021 (Finalist)
EY
-
-
EY: Entrepreneur of the year 2021 (Finalist)
EY
-
Deloitte Technology Fast 50 Award
-
-
Great Place to Work
-
(Best employer in Münsterland, in the entire IT industry and Germany)
-
Deloitte Technology Fast 50 Award
-
-
Great Place to Work
-
(Best employer in Münsterland, in the entire ITC industry and Germany)
-
Innovator des Jahres
Brand Eins
-
Deloitte Fast 50 Award
-
-
Deloitte Sustained Excellence Award
-
-
Great Place to Work
-
(Best employer in Münsterland, in the entire ITC industry and Germany)
-
Deloitte Technology Fast 500 Award (worldwide)
-
-
Great Place to Work
-
(Best employer in Münsterland, in the entire ITC industry and Germany)
Other similar profiles
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top contentOthers named Stefan Hamann in Germany
24 others named Stefan Hamann in Germany are on LinkedIn
See others named Stefan Hamann