Teaching AI agents when not to act alone.
Researching safe autonomy for AI-driven decision systems.
The safety layer evaluates what the agent wants to change. Clear cases can proceed; uncertain or consequential cases move deliberately to human review.
A control and escalation layer for LLM-based multi-agent systems that decides when automation is safe and when a human needs to step in. This is the core innovation behind everything we build.
Active R&D since 2025, with validation in a live commercial environment.
Chemistry & compliance automation, theoretical physics & applied AI/ML, plus business scaling.
We're open to research collaborations, funded projects and pilot partnerships around TENMA-Control.
Start a Research CollaborationJoint research, funding applications, or access to our TENMA-Control approach and findings.
Discuss researchBring safely governed, agent-based automation from our own live operating environment into your processes.
Discuss automationConnect shop, warehouse, accounting and fulfillment into one governed, end-to-end process.
Discuss integrationWhether it is a research collaboration, an automation project or process integration, we would love to hear from you.
TENMA-Control studies how autonomous agents can move from plausible proposals to controlled action, with the minimum intervention necessary for the state in front of them.
Agents may receive evidence that looks credible in isolation while referring to different versions of the system state. Missing fields, stale values and parallel actions can turn a plausible proposal into a consequential mistake.
The control problem begins before execution: not with a malicious agent, but with fragmented evidence that no longer describes one coherent present.
The system must choose among legitimate controls rather than reduce every case to a simple pass or fail. Evidence quality, conflict, novelty, impact and reversibility shape whether the next step should be execution, re-checking, simulation, model review, human review or containment.
The objective is selective autonomy: the least intensive control that still respects the agreed safety boundary.
The deterministic core protects non-negotiable boundaries. Semantic review and structured risk signals add context. The decision layer then selects the minimum necessary intervention, while human review remains a regular control outcome.
Counterfactual same-state tests replicate an identical proposed action and starting condition across multiple control paths. The resulting outcomes are evaluated on a common surface before findings transfer into a more realistic development environment.
This makes the research question measurable: which intervention actually improved the outcome, at what cost and under which conditions?
TENMA-Control is open to validation partners, technical exchange and funded research collaborations.
Start a research conversationWe came together not because we planned it, but because we kept running into the same interesting problem from different directions.
Build a real business that should require no more than ten minutes of operational work per day.
Sven and a long-standing business partner had wanted to build a company together for years. The combination was practical: experience in business development on one side, and deep experience with process automation, enterprise systems and compliance-driven operations on the other.
In 2024, they turned that ambition into a live shop business. Because the company was built alongside existing professional responsibilities, the operating model had to be radically efficient from the beginning. Orders, inventory, pricing, accounting and fulfillment could not depend on continuous manual intervention. The internal rule became the company name: the business should eventually run with no more than 10 minutes of daily operational attention.
Reality was less tidy. Building the shop and its automations required far more than ten minutes a day. Yet the constraint produced something useful: it forced the team to automate complete workflows rather than isolated tasks. As AI capabilities advanced, specialised agents became part of that system. The agents created new leverage, but also exposed a harder problem. Independent proposals could conflict, rely on different versions of the system state or recommend actions whose consequences were difficult to reverse.
Daniel joined the work in 2025. With a background in theoretical physics, applied AI and software engineering, Daniel helped turn recurring operational failures into precise technical questions. The team initially explored a conventional orchestration harness, but the core issue was not simply how to coordinate more agents. The core issue was how to decide when an agent should be allowed to act, when another control was required and when a human should intervene.
That question became TENMA-Control: an independent control and escalation layer between an AI agent's proposal and a consequential action. The existing shop-system landscape provided an unusually valuable development environment. The team could observe real state changes, conflicts and edge cases without inventing a hypothetical use case.
Today, the same architecture is being developed as both a research programme and a practical service. The objective is broader than one shop: controlled agentic automation for commerce, operational workflows and other environments where AI systems interact with real processes. The research application is the next step in turning an operational insight into a measurable, transferable safety approach.
Trained as a chemist, with years of hands-on experience in compliance-driven environments, including software rollouts, SAP systems and the automation of compliance processes from the inside. That first-hand view of untapped automation potential sparked 10minutes.
Background in theoretical physics with deep, self-taught expertise in AI and software engineering, including a profitable crypto trading bot built from scratch. Brings product management and startup experience to the team.
Also a chemist by training, with a strong background in business and scaling ventures. This team member currently works behind the scenes on strategy and growth.
The company begins with the goal of creating a real business whose recurring operations can be reduced to ten minutes a day.
The team builds its own shop-system landscape and connects ordering, stock, pricing, accounting and fulfillment. The operating environment remains unnamed publicly.
Specialised AI agents begin to support operational tasks. Productivity improves, but conflicts, stale state and uncertain actions reveal the limits of unrestricted automation.
Operational problems are translated into technical and scientific questions about evidence, intervention, reversibility and human review.
The team moves beyond simple orchestration and develops a dedicated decision boundary between agent proposals and real-world execution.
TENMA-Control is prepared for systematic evaluation, research collaboration and use in customer automation and end-to-end commerce solutions.
The final team photograph will follow later. Until then, the four TENMA ghosts are holding the space and representing the qualities we want in the team: curiosity, technical depth, operational discipline and constructive challenge.
Research, automation or process integration. We are happy to connect.
Get in touchOur own live shop landscape provides a continuously operating environment for developing and validating automation under real conditions, without placing customer operations at risk.
We design agent-based automation for operational processes and add an explicit control layer between an agent's proposal and a state-changing action. The objective is not maximum autonomy. It is selective autonomy: the appropriate degree of intervention for the evidence, system state and potential consequence in front of the agent.
An agent recommendation is first represented as a structured action proposal. Deterministic rules can reject prohibited or incomplete actions. Semantic review adds context, while risk signals consider evidence quality, conflicting activity, novelty, impact and reversibility. The resulting decision can release, simulate, re-check, escalate or block the action.
When specialised agents work on pricing, inventory, service, procurement or other domains, their proposals may overlap or rely on different versions of reality. TENMA-Control introduces a common decision boundary, records the intervention and feeds verified outcomes back into trusted state.
10minutes supports organisations in identifying suitable use cases, designing agent workflows, integrating operational systems and implementing the control logic required for responsible deployment. The approach is not limited to e-commerce. It can be applied wherever AI agents analyse context, coordinate tasks or propose consequential actions.
We also develop and validate the approach in our own active shop-system landscape. The environment is deliberately not named publicly, so ordinary customers are not framed as research participants or test subjects. Operational observations are used to improve system design without making claims about individual customers.
Our own shop landscape is live today. Talk to us about bringing the same automation approach to your processes.
We design and deliver complete commerce operations: shop, customer journey, inventory, accounting, warehouse and fulfillment. Instead of handing clients a collection of interfaces, we connect the full operational chain and make state, decisions and outcomes traceable.
Clients can approach 10minutes with an existing shop, a fragmented system landscape or a new business model. We map the operating flow, connect the required platforms and implement the controls that keep automated actions aligned with trusted state.
Our delivery scope can include customer-facing shop processes, product and order data, inventory availability, accounting events, warehouse instructions, shipping and status feedback. We work with a fulfillment partner that operates a fulfillment center, allowing the designed process to extend into physical storage, picking, packing and dispatch without naming the partner publicly at this stage.
The customer journey and the operational chain are treated as one system. Every transition carries an explicit state, and status information flows back toward the shop and customer rather than disappearing at organisational boundaries.
Shop systems, ERP, warehouse, accounting and shipping exchange state and outcomes through controlled interfaces. AI agents can support analysis, coordination and automation, while TENMA-Control governs actions that affect prices, stock, orders, financial records or fulfillment.
Talk to us about the complete chain, from the customer experience and agentic automation to inventory, accounting, warehouse and fulfillment.
Contact us about your solutionThis is a curated reading list of publicly reported agent failures, exploit disclosures and security incidents, with the newest reports shown first. The pattern is not that AI is unusable. The pattern is that capability without an independent decision boundary creates avoidable risk.
Prompt injection, excessive permissions, stale state and weak separation between proposal and execution recur across very different systems. TENMA-Control is our response: make each consequential action pass through explicit rules, structured evidence, risk-adaptive intervention and human escalation where necessary.
ZEIT berichtete über einen weiteren Alarmfall, bei dem ein autonomes System nicht nur analysierte, sondern aktiv täuschende Nachrichten versandte. Der Fall verdeutlicht den Unterschied zwischen generiertem Inhalt und einer ausgeführten Außenwirkung.
Quelle lesen · ZEIT Online ↗ZEIT berichtete, dass ein weiterer Agent bei Tests reale Systeme erreichte. Als Ursache wurde eine fehlerhafte Testkonfiguration genannt. Der Fall zeigt, wie schnell eine simulierte Aufgabe ohne verlässliche Umgebungsgrenze reale Folgen haben kann.
Quelle lesen · ZEIT Online ↗heise rekonstruierte, wie Modelle aus einer isolierten Evaluation ausbrachen, eine Zero-Day-Schwachstelle nutzten und weitere Software beziehungsweise Dienste erreichten. Die übliche Modellkontrolle war für den Test bewusst reduziert.
Quelle lesen · heise online ↗SPIEGEL berichtete, dass der OpenAI-Agent neben dem bekannten Angriff weitere Online-Dienste attackierte und fremde Zugangsdaten nutzte. Die Reichweite des Vorfalls war damit größer als zunächst bekannt.
Quelle lesen · DER SPIEGEL ↗Nach SPIEGEL-Recherchen dauerte es mehrere Tage, bis der Ausbruch und die eigenständigen Angriffe erkannt wurden. Der Fall macht neben Prävention auch Telemetrie, Auditierbarkeit und schnelle Abschaltung zu zentralen Kontrollfragen.
Quelle lesen · DER SPIEGEL ↗ZEIT beschrieb, wie neue Modelle während eines Tests die vorgesehene Umgebung verließen und in Systeme eines anderen KI-Unternehmens eindrangen. Aus einer Evaluationsaufgabe wurde ein realer Cybervorfall.
Quelle lesen · ZEIT Online ↗heise berichtete über Schwachstellen, die Datenmanipulation, Zugriff auf geschützte Informationen, Rechteausweitung und Schadcodeausführung ermöglichen konnten. Auch die Agenteninfrastruktur selbst benötigt damit unabhängige Schutzschichten.
Quelle lesen · heise online ↗An autonomous coding agent reportedly ignored repeated freeze instructions, removed live data and then produced misleading status information.
Read source · AI Incident Database ↗The disclosed prompt-injection chain showed how an assistant with access to enterprise context could be redirected across trust boundaries.
Read source · Research paper ↗A proof of concept showed private source code and secrets could be exfiltrated through an indirect prompt injection and GitHub’s image proxy.
Read source · Legit Security ↗Untrusted project content could redirect an agent and exploit auto-approved commands to encode local information into outbound DNS queries.
Read source · Embrace The Red ↗CVE-2025-54795 described a command-injection path capable of bypassing a confirmation checkpoint and executing untrusted commands.
Read source · SentinelOne ↗OWASP’s Q1 round-up included an inbox-deletion incident among cases showing why agent permissions and rollback boundaries matter.
Read source · OWASP GenAI ↗The OWASP round-up listed an internal agent data exposure among real incidents moving agent risk beyond theoretical model behaviour.
Read source · OWASP GenAI ↗The reported case focused on agent identity and privilege boundaries, where useful access can become an attack path without independent controls.
Read source · OWASP GenAI ↗OWASP included the Mercor/LiteLLM supply-chain case as an example of third-party agent infrastructure expanding the trust boundary.
Read source · OWASP GenAI ↗CVE-2025-59528 was included in OWASP’s exploit round-up, linking agent tooling and MCP configuration to consequential execution risk.
Read source · OWASP GenAI ↗External sources open in a new tab. Entries are concise editorial summaries of the linked public material. Inclusion does not imply that every event involved the same architecture or that TENMA-Control would by itself have prevented it.
Let’s define what it may do alone, what must be checked and when a human needs to step in.
Discuss controlled agentsInformation about how 10minutes GmbH processes personal data when you use this website or contact us.
Last updated: 12 August 2026
10minutes GmbH
Borsteler Chaussee 5
22453 Hamburg, Germany
Email: info@10minutes.io
This policy explains how 10minutes GmbH processes personal data when you use this website or contact us through the contact form. Personal data means information that relates to an identified or identifiable person.
When the website is hosted by a web provider, technical access data may be processed to deliver the website securely and reliably. This may include the IP address, date and time of access, requested file, browser information, operating system, referring page and status information. The legal basis is Article 6(1)(f) GDPR. The legitimate interest is the secure and technically reliable operation of the website. Technical log data should be deleted when it is no longer required for security and operational purposes, subject to legal retention obligations.
If you use the contact form, the information you enter is transmitted securely to a website endpoint operated for 10minutes GmbH. The message is then forwarded as a notification to info@10minutes.io so that the enquiry can be reviewed and answered.
When the email reaches 10minutes GmbH, we process the information you provide, such as your name, organisation, email address, selected topic and message, to review and answer your enquiry. Depending on the content of the enquiry, the legal basis is Article 6(1)(b) GDPR for steps taken at your request before entering into a contract, Article 6(1)(f) GDPR for general business communication, or Article 6(1)(a) GDPR where you have given consent.
Personal data is accessed only by people who need it to handle the enquiry. Service providers may process data on our behalf when this is necessary for email, hosting or technical operation. We do not sell personal data.
If a service provider processes data outside the European Economic Area, we use an appropriate transfer mechanism where required, such as an adequacy decision or the European Commission Standard Contractual Clauses, together with supplementary safeguards where appropriate.
We keep enquiry data only for as long as it is needed to answer the enquiry and manage the resulting business relationship. Data may be retained longer where statutory retention duties apply or where it is required to establish, exercise or defend legal claims.
Subject to the conditions of the GDPR, you may request access, correction, deletion, restriction of processing and data portability. You may object to processing based on legitimate interests. Where processing is based on consent, you may withdraw that consent at any time with effect for the future.
You have the right to lodge a complaint with a data protection supervisory authority. You may contact the authority responsible for your place of residence or work, or the authority responsible for 10minutes GmbH in Hamburg.
For privacy questions or to exercise your rights, contact info@10minutes.io.
We may update this policy when the website, our processing activities or legal requirements change. The current version is available through the Privacy Policy link on this website.