Madison Faulkner

The looming collapse of human-centric pipelines Traditional software development relies on a predictable, slow-moving rhythm. A human developer spends hours or days crafting a pull request (PR), which then sits in a queue for a colleague to review. This CI/CD model operates on the assumption that code changes are discrete, occasional, and manageable by human eyes. Madison Faulkner from NEA points out that while a human might submit one or two diffs a week, autonomous agents operate at a scale that breaks this fundamental premise. When thousands of short-lived branches pull a codebase in different directions simultaneously, the result is infrastructure chaos. Runner saturation, cache thrash, and cold Docker builds become the norm rather than the exception. The merge queue, once a simple gatekeeper, transforms into a high-performance database problem requiring serialized locks that choke development velocity. Moving from PRs to intent-based loops Hugo Santos, CEO of Namespace, argues that the industry must move beyond the pull request entirely. In an agentic world, development starts with a codified intent—a specification or a ticket—rather than a code change. This intent enters a stateful agent loop where the AI checks out a well-known commit and iterates toward the goal using internal validation assets. This shift moves evaluation into the "inner loop." Instead of waiting 15 to 45 minutes for a CI server to report a failure, the agent performs continuous, stateful validation. This environment preserves memory and state, avoiding the inefficiency of starting from scratch with every iteration. The goal is to eliminate the latency currently hidden behind human work patterns. Pre-merge queues and the multiverse of commits The proposed architecture replaces traditional gates with a pre-merge queue. Because agents generate code faster than humans can vet it, the human role shifts from reviewing line-by-line diffs to approving intent-plus-outcome. A reviewer might watch a video of a feature working or examine a security-focused LLM's report rather than parsing thousands of lines of syntax. In the near future, development may enter a "multiverse" state. Agents will explore multiple potential commits in parallel for the same plan, navigating a repository where the "tip" is a moving target. This requires massive compute capacity and a fundamental rethinking of how we maintain serializability in a ledger of code changes. The evolution of governance and CI CI is not disappearing; it is being absorbed into the harness itself. Governance moves from a separate phase to a continuous constraint. The harness coerces changes to follow team standards in real-time, ensuring that compliance and invariants are enforced with every iteration of the agent. This transition represents a shift from static pipelines to continuous compute environments that treat code as a fluid, ever-evolving asset.