CI/CD Is Dead, Agents Need Continuous Compute and Computers — Hugo Santos and Madison Faulkner

Madison Faulkner (partner at NEA, formerly Meta AI researcher) and Hugo Santos (CEO of Namespace, formerly Google microservices lead) argue that traditional CI/CD pipelines are fundamentally incompatible with agentic software development at scale, and lay out a concrete replacement architecture they are already deploying with customers.

The core problem is one of throughput mismatch. Traditional CI/CD was designed around human developers submitting one or two diffs per week. At agent scale, thousands of autonomous agents open PRs continuously against the same codebase, causing runner saturation, cold Docker builds on every short-lived branch, cache thrash, and merge queues that begin to behave like serialized database locks — where time-to-commit becomes the primary bottleneck rather than the time to write code. Santos illustrates this with real-time GitHub commit activity data showing a dramatic spike in commits and lines added/deleted in recent months, driven largely by agents.

Faulkner frames the transition as a shift from monolithic agents (a single LLM engine) to microservices-with-agents — a more distributed, composable model of autonomous development. She highlights Mitchell Hashimoto's (HashiCorp founder) public prescription for fixing GitHub as validation: shut down Copilot-style bolt-ons and instead rebuild around cloud-native, inference-aware infrastructure that serves AI and agentic users first.

The proposed replacement architecture eliminates PRs entirely as the unit of work. Instead, the flow begins with a codified intent-and-plan (captured in a Linear ticket, Slack thread, or spec document) fed into an agent harness (examples cited: Claude Code, Amp, Cursor, Factory). Within the loop, the agent checks out a well-known commit, implements the plan, runs internal validation (build + tests), then surfaces a human-readable outcome — not a diff — for lightweight human confirmation ('continue'). The validation pipeline is inline to the loop rather than a post-hoc gate.

For external validation, the vision moving from months to weeks is to replace human code reviewers with specialized LLM agents: a security-focused LLM, an API-conformance LLM, and others that provide structured feedback directly back into the main agent harness. This requires stateful execution environments — starting from scratch on every iteration is too slow and erodes the performance gains from faster inference.

Before changes reach the main repository, they flow into a pre-merge queue — a reconciliation layer that handles serializability across many concurrent agent branches touching the same code. Human review at this stage is on intent-plus-outcome (including feature demo videos or LLM security reports), not on raw diffs. Namespace reports their own team's PR-equivalent volume is already 4x what it was historically, making per-PR human review untenable.

Looking further ahead, Santos describes a 'multiverse' model where the tip of the main branch is itself a moving target. Agents may explore multiple parallel commits simultaneously for the same plan, choosing the best candidate before merging. This multiplies resource consumption substantially and makes the performance and statefulness of the inner loop critical infrastructure rather than a convenience. The speakers emphasize that CI as a concept does not disappear — invariants like starting from a vetted, well-known checkout remain essential for compliance — but CI shifts from a separate pipeline phase to a continuous, embedded property of every iteration of the agent loop. Governance also moves up into the harness itself rather than being enforced externally.