SmartBear CEO Dan Faulkner on the quality gap created by agentic coding, why intent validation is quality's next big role, and what BearQ is trying to do about it.

The growing disparity between the speed at which code is being produced, accelerated by AI coding tools, and the capacity of quality, security, and governance activities to keep pace with that output. As the inner loop accelerates and more software is generated faster, the outer loop becomes a bottleneck. The quality gap refers to this imbalance: more software arriving at testing and review stages than teams have the time or tools to handle thoroughly, creating pressure to either sacrifice coverage, slow down releases, or find new ways to scale quality work.

The set of SDLC activities that happen after a developer's local coding work is complete, including integration, code review, quality assurance, security testing, governance, and deployment. The outer loop is contrasted with the inner loop, which covers the individual developer's fast local cycle. When AI tools accelerate the inner loop, the outer loop often becomes the bottleneck, because the teams and tools responsible for those downstream activities have not sped up to match. Quality, security, and DevOps disciplines all live primarily in the outer loop.

A structured map of everything an application does — all its pages, pathways, workflows, and API calls — built by systematically exploring the application rather than reading its source code. In an agentic testing context, a knowledge graph is generated automatically by an agent that navigates the running application and records what it finds. It gives teams a ground-truth picture of what has actually been built, which may differ from what was specified. It can be used to inform test strategy, surface coverage gaps, and support intent validation by comparing the graph against original requirements.

A framework for describing levels of automation in software development, adapted from the levels of autonomy used in autonomous vehicle design. At the lowest level, all work is done manually by a person. At each higher level, more decision-making and execution is handed to tooling or agents, with the highest level representing fully autonomous operation requiring no human intervention. The framework is used to help teams understand where they currently sit, where they want to get to, and how to move between levels safely. In a testing context, the ladder might run from manual test creation, through low-code automation, through AI-assisted test generation, up to a fully agentic testing system that explores, writes, executes, and reports without human instruction.

The rate at which software is successfully delivered into production, as distinct from coding velocity. A team may write code faster due to AI tools while production velocity stays the same or even drops, because quality, security, and governance activities downstream have not accelerated to match. Production velocity is the more meaningful measure of a team's actual throughput, since it reflects the full system of constraints in the SDLC rather than just the speed of the coding step.

The fast, localised development cycle covering writing, running, and testing code on a developer's own machine, before changes are pushed to shared infrastructure. The term is used to distinguish the individual developer's coding and unit testing step from the broader SDLC activities that follow, sometimes called the outer loop, which include integration, review, quality assurance, security, and deployment. In the context of AI-accelerated development, the inner loop has sped up significantly, creating pressure on the outer loop to keep pace.

The process of checking whether an agentic coding tool actually built what was asked for, and whether it built anything that was not asked for. As AI coding agents have latitude to make their own decisions, the output they produce can drift from the original specification in ways that are not immediately visible. Intent validation sits upstream of functional testing: before asking whether the software works, it asks whether the software is the right thing. It requires a combination of system knowledge, user empathy, and judgement about what constitutes acceptable drift versus what must be sent back. For example: reviewing a knowledge graph of what an agent built against the original user stories; checking for features added without being requested; or confirming that access and permission models reflect what was specified rather than what the agent assumed.