Web Summit returns to Lisbon on 9–12 November 2026. Its published programme puts three questions close together: who controls AI agents on the web, what open models offer builders, and how much of AI research AI can perform. This is a preview of planned conversations—not a report of products launched or breakthroughs proven at the event.
Quick Take
- Confirmed programme: Web Summit lists “When AI builds itself,” “The open-source advantage” and “The agentic internet: Who controls the web?” for 10 November. The schedule is being updated as the event approaches.
- Agentic web: The practical test is whether an agent can identify itself, respect site controls, show the user what it will do and act with limited, revocable permission.
- Open models: Judge the rights to use and modify a model, the information available to audit it, and the cost of operating it—not the label “open” alone.
- AI building AI: A research agent can propose and run bounded experiments without becoming an unrestricted system that designs and releases its own successor. Ask for reproducible results and human review.
Why Does Web Summit 2026 Matter for AI?
Web Summit says its 2026 Lisbon event runs from 9 to 12 November. Its published schedule already includes the three headline conversations on 10 November, and the organiser says it is adding more sessions. The programme reflects questions the industry wants to debate. It cannot tell us in advance what speakers will demonstrate, disclose or conclude.
The themes are connected by control. An agent acting for a person must know which sites it may use. A team choosing an open model must know which parts it can inspect and run. A research lab allowing AI to improve code or experiments must decide which changes require human approval. All three promise more autonomy, but each also moves decisions into systems that are harder to observe than a conventional chat window.
Web Summit’s AI Summit track also highlights agentic systems, open versus closed approaches and the return on AI investments. That makes the event useful for MaGeN-AI readers who want to move beyond announcements: what works in a deployed workflow, what it costs, who can inspect it and who is accountable when it fails?
1. The Agentic Internet: Who Gets to Act on the Web?
The scheduled “The agentic internet: Who controls the web?” session lists Cloudflare co-founder and CEO Matthew Prince and The Atlantic CEO Nick Thompson. Their roles make the question concrete for both infrastructure providers and publishers: when software reads a page, compares options and completes a transaction for someone, how can a site tell an authorized customer agent from unwanted automated traffic?
Imagine asking an assistant to book a hotel. It needs the right dates, a spending limit, access to a booking service and perhaps a payment method. A hotel site needs to know whether the request comes from a genuine customer, whether its listing is being copied for another purpose and who handles an error. The customer needs a clear chance to check room terms and the final price before a purchase. In that workflow, identity, permission and confirmation matter as much as the model’s ability to browse.
There is already a standard for expressing crawler preferences: the Robots Exclusion Protocol. Its specification explicitly says these rules are not access authorization. An agent visiting a public product page, signing into an account and buying something crosses different boundaries. A useful system would respect crawling rules where applicable, use proper authentication for protected services and require explicit authority for consequential actions.
Listen for practical answers on rate limits, agent identification, user consent, attribution to original sources, compensation for publishers and dispute handling after an erroneous transaction. A live browser demo is interesting; a repeatable method for respecting those boundaries is more important. These are questions for the session, not capabilities Web Summit has promised to announce.
2. The Open-Source Advantage: Open for Whom, and for What?
Web Summit’s “The open-source advantage” is listed with Genspark co-founder and CEO Eric Jing, ModelBest co-founder and CEO Dahai Li, and True Ventures partner Rohit Sharma. Its programme description presents open AI as a route to control over data, lower compute costs and faster development. Those are propositions to examine, not outcomes guaranteed by choosing an open model.
A downloadable weight file can let an organization run inference on its chosen infrastructure. But the term open source AI carries a broader expectation. The Open Source Initiative’s definition discusses freedoms to use, study, modify and share, along with access to appropriate code, parameters and information about training data. A license that permits a demo may impose conditions on production use. A model that exposes weights may still provide limited visibility into training.
For an enterprise, ask what exactly is available: model weights, training and inference code, a usable license, evaluation methods and any required tooling. Then count the full operating bill—hardware, serving, monitoring, updates, security review and people who maintain it. Running a model yourself may improve control over deployment and data paths. It does not automatically make inference cheaper or a system safer.
The most useful comparison is a cost per successfully completed task at an acceptable quality level. An inexpensive model that needs frequent retries or human correction can cost more than an API with a higher token price. Conversely, an appropriately sized model running under an organization’s control may be valuable where auditability or predictable deployment is essential. The event should give builders enough detail to test those trade-offs.
3. When AI Builds Itself: Research Assistance or Autonomous Improvement?
The “When AI builds itself” programme lists Recursive and You.com co-founder and CEO Richard Socher with Wired senior writer Kate Knibbs. The title poses a striking possibility, but it is not itself evidence that an AI system independently developed and safely released its successor.
There are several distinct levels of automation. A coding assistant can suggest a better training loop or fix a test. A bounded research agent can search prior experiments, propose an architecture change, run approved evaluations and write up the findings. A system that selects its own objective, changes its own evaluation rules, allocates compute and releases a new model without human checks would be a much stronger claim. Conflating those levels makes both progress and risk difficult to assess.
A credible case study should show an initial human research question, what data and tools the agent could access, which candidate changes it produced, and the measurements used to accept or reject them. Teams should separate gains in research throughput from gains in model quality. Running more experiments per week can be valuable even if most proposals fail. Claims of a better model need a reproducible evaluation on tasks the agent could not simply optimize by seeing the answers.
Safety also becomes a workflow issue. Who can approve a change to training data, an evaluation set or deployment policy? Are failed runs preserved for review? Can a researcher stop or roll back a long-running experiment? Those questions matter more than a dramatic headline about self-improvement, especially if thousands of agents or jobs contribute small pieces to a larger research programme.
Where Do the Three Themes Meet? Durable, Portable Infrastructure
A fourth scheduled session, “Bet on the foundation, not the framework,” features Temporal co-founder and CEO Samar Abbas. The organiser’s description says the talk will argue that the infrastructure executing a workflow matters when agent frameworks and models change. That is the speaker’s proposed argument; it should be tested against actual system designs.
The connection is practical. A travel agent may need to pause for a person to approve a purchase. An enterprise using an open model may switch models while keeping a reliable transaction log. A research agent may run an experiment overnight and need to resume from a known state after an interruption. All three benefit from explicit permissions, checkpoints, traceability and recovery when tools fail.
When a conference demo claims to be “agentic,” ask what happens after a timeout, a changed API response or a denied action. If the answer depends on a human silently restarting the task, the polished final screen hides a significant part of the deployment cost.
Web Summit AI Sessions: What Should Readers Watch For?
The sessions below are listed on Web Summit’s programme for 10 November 2026. Titles, speakers and timings can change; check each linked session page shortly before the event.
| Scheduled session | Central question | Evidence worth seeking |
|---|---|---|
| When AI builds itself | Which research tasks can agents perform with human supervision? | Before-and-after results, independent evaluation, experiment logs and an explicit human approval boundary. |
| The open-source advantage | What control does openness actually give builders? | Specific artifacts and license terms, quality at a stated workload, and total deployment cost. |
| The agentic internet: Who controls the web? | How do sites, users and agents share authority? | Agent identity, site controls, user authorization, provenance and transaction safeguards. |
| Bet on the foundation, not the framework | Can workflows survive model and framework changes? | Recovery from failed tools, portable logs, replayable steps and a working human approval gate. |
Who Should Pay Attention to These Discussions?
Watch how agent traffic is identified, how answers credit original work and whether content owners can express access and licensing preferences that agents actually respect.
Ask how an agent proves customer authority, handles price or stock changes, requests final purchase approval and recovers when a booking partly succeeds.
Compare model portability, license conditions, logs, durable execution and the cost of human review. A working prototype is the start of the test, not the end.
Look for bounded tasks, reproducible evaluations, data-access controls, clear ownership of decisions and procedures for stopping an agent or rejecting a proposed change.
How Should AI Claims From Web Summit Be Checked?
- Identify the type of claim: A session title, speaker forecast, product roadmap, live demo and independent research result carry different levels of evidence.
- Ask for a baseline: If a speaker cites a speedup, cost saving or improved accuracy, what existing workflow was measured, for how long and on which tasks?
- Include the complete workflow: Track tool failures, model retries, human approvals, corrected errors and time to an accepted result.
- Check who can repeat it: Are the evaluation criteria, usage terms, deployment requirements and sample tasks sufficiently clear to assess the claim elsewhere?
- Verify after the event: Link any new product assertion to its own announcement or technical documentation and date the article update. A talk about a future system is not proof it is available now.
What About Copyright, Privacy and the EU AI Act?
Automated web use raises questions about rights to access and reuse material. A site being publicly viewable does not settle what may be copied, republished, used for training or used to answer users without attribution. Builders should check applicable permissions and licenses for text, images, datasets and model components. This article is original commentary; it links to the organiser’s programme without copying session descriptions, graphics or speaker photographs.
Agents handling personal information should collect only what a task needs, limit who receives it and make consequential actions reviewable. The European Commission’s GDPR guidance describes principles such as purpose limitation and data minimization. An agent booking a hotel and an agent evaluating a job applicant may have very different data and legal implications.
The European Commission’s AI Act guidance explains that applicable obligations depend on the actor’s role and the system’s intended use. An open model or a web agent is not automatically classified as high-risk merely because it is powerful. A tool intended for certain employment or credit decisions may trigger different requirements from a general information assistant. Conference claims of safety or compliance should be checked against the specific deployment.
MaGeN-AI View
The common question is who can say yes—and who can say stop
Web Summit 2026’s AI agenda captures a real shift in focus: from a model answering prompts to systems that browse, run tools and help build future systems. The most useful announcement will be one that shows clear authority, measurable outcomes and a way to recover from mistakes. Until the talks happen, these sessions are a guide to the questions MaGeN-AI will ask, not evidence that the industry has solved them.
Frequently Asked Questions
When and where is Web Summit 2026?
Web Summit 2026 is scheduled for 9–12 November in Lisbon, Portugal. Its programme may change as the event approaches.
Which Web Summit 2026 AI sessions does this preview cover?
The published programme lists “When AI builds itself,” “The open-source advantage” and “The agentic internet: Who controls the web?” for 10 November. The article also discusses “Bet on the foundation, not the framework.”
What does the agentic internet mean?
It refers to web experiences in which AI agents find information or use services on a person’s behalf. Important questions include site access, agent identity, user permission and accountability for actions.
Are open-weight AI models automatically open source?
No. Downloadable weights alone do not establish the permissions and supporting code, parameters and training-data information discussed in the Open Source Initiative’s Open Source AI Definition.
Does “When AI builds itself” prove AI can develop its own successor?
No. It is the title of a scheduled discussion. Assess any later claim by checking the tasks agents actually performed, the evaluations used and the human approvals involved.
