CI/CD Dies at Agent Scale: Agent Loops + Continuous Compute

The Breakthrough

Madison Faulkner and Hugo Santos propose replacing CI/CD with continuous compute in agent loops: codify intent and plan, run fast stateful validation inline without PRs, use premerge queues for serialization, and have humans approve intent-plus-outcome bundles from parallel agent work.

What Actually Worked

• Codify intent and plan as a spec in Linear tickets or Slack messages; feed it into an agent harness like AMP, Cursor, or Factory.
• Agent checks out a well-known commit, implements the plan incrementally, and performs internal validation by building and testing against repository assets.
• Human intervenes briefly to approve outcomes or say 'continue,' looping until plan completion before entering the merge queue.
• Shift validation into the inner loop with other agents (e.g., security-focused LLM, API conformance LLM) providing fast feedback without humans.
• Use a premerge queue to serialize parallel changes semantically, grouping them for human review of results like feature videos or LLM outputs, not code diffs.

Context

Traditional CI/CD handles human-scale diffs (one or two per week) with PRs, reviews, builds, tests, and merges, but agents generate thousands of short-lived branches, causing runner saturation, cold Docker starts, cache thrashing, and merge queues acting like serialized database locks.

Leading teams (e.g., FAL, Zed, RAMP, Namespace) already skip PRs for continuous work: humans previously hid machine slowness, but accelerating code gen forces validation into fast inner loops. Future multiverse explores multiple repo commits in parallel per plan, demanding stateful environments, world signals for plan adaptation, and massive compute capacity.

Notable Quotes

• "Continue is probably the word that we use the most nowadays."
• "The act of merging um is starting to look a lot like um high performance uh uh database problems where you have serialization and you have a single ledger."
• "Agents may actually be working on multiple commits at the same time to address the same plan... this inner loop needs to be extremely quick."

Content References

Namespace accelerates this via hardware/software co-design for cache orchestration, ingress shaping, rate limiting, and agentic identity.