Your AI Agent Wrote the Code. Who Verifies It?
How to build closed-loop verification for agentic software development
Agents now write code faster than anyone can verify it.
Agents can write features, open PRs, and ship code at a pace traditional QA workflows weren't designed to handle. Prompts, specs, and rules files help, but they're still open-loop control: nothing actually tells the agent when it broke the product.
So verification lands back on the developers writing and reviewing the code: reading diffs, clicking through user journeys by hand, keeping scripted suites alive. Having a QA function helps, but it doesn't change the arithmetic. A single agent can produce more change in an afternoon than a review process built for human output absorbs in a week.
What happens when the agents writing the code can also get real feedback on whether it actually works?
Join Vilhelm von Ehrenheim, Co-founder & Chief AI Officer at QA.tech, for a technical deep dive into how autonomous agents can close the loop between AI-generated code and production-ready software.
What We'll Cover
You'll learn:
Why better prompts and specifications won't close the loop, and why verifiable feedback matters more the more of your codebase an agent writes.
What's actually under the hood of an autonomous verification agent: knowledge graphs, world models, visual browser agents, hierarchical context, and perception-action loops.
Why goal-based testing behaves differently from scripted automation, and how an agent verifies that a user got what they came for rather than that a selector still resolves.
How dynamic PR testing works, from reading a code diff to identifying affected user journeys, running tests against a preview environment, and returning evidence directly to the PR.
How to rethink regression around a small critical-path suite plus dynamic testing, change-scoped verification, so coverage stops being a maintenance project.
Live Demo: From PR → Verification
Vilhelm will be running the full workflow live:
PR opens → the diff is analyzed → affected journeys are identified → tests are generated → agents execute them against a preview environment → evidence and a verdict return to the PR.
You'll see how the agent builds and uses its knowledge of the application, how tests get generated from goals, an agent working through a real app on its own, and the screenshots, logs, recordings, and reasoning it produces when something fails.
Who Should Attend
This session is built for CTOs and engineering leaders, the AI and software engineers doing the work, and the QA and product people who own quality alongside them, on teams already shipping with agentic development.
Especially if your team is reaching the point where:
Agents can open PRs faster than humans can verify them.
About Vilhelm
Vilhelm von Ehrenheim is Co-founder and Chief AI Officer at QA.tech, an AI-native end-to-end testing and verification platform based in Stockholm.
Before QA.tech, Vilhelm spent five years building the Motherbrain AI platform at EQT, one of the earliest production deployments of machine learning in venture capital. Prior to EQT, he led the Predictive Modeling team at Klarna, building real-time credit and fraud risk models in a decision pipeline processing roughly 300,000 transactions per day.
His applied research has been published at EMNLP, KDD, and CIKM, and he holds an MSc in Engineering Physics from Lund University.
—
Format: ~35-minute technical talk + live demo, followed by Q&A.
Hosted by AlphaSignal read by over 300,000 AI engineers, researchers, and technical founders.
Subscribe today at AlphaSignal.ai