← back
CI/CD Is Dead, Agents Need Continuous Compute and Computers — Hugo Santos and Madison Faulkner
Takeaway
Agent-driven development needs a continuous-compute substrate that replaces PR-centric CI/CD with high-throughput, machine-paced merging.
Summary
- NEA's Madison Faulkner and Namespace's Hugo Santos argue traditional CI/CD breaks under agent-scale PR volume; merging across thousands of short-lived agent branches becomes a database-locking problem.
- Proposed 'continuous compute' layer: hardware/software co-design with intake shaping, orchestrating cache, agentic identity, retries at scale, sitting over GitHub Actions.
- New unit of work isn't the PR but intent+plan+spec feeding an agent harness (Amp, Claude Code, Cursor, Factory) that checks out known commits with built-in invariants.
- Cite Mitchell Hashimoto's GitHub redesign and customers like Fal, Zed, Ramp pioneering the pattern; GitHub commit/line-change spikes already visible in the wild.
ci-cdcoding-agentsinfrastructure
Original description
Traditional CI/CD was built for humans pushing one or two diffs a week. Scale to thousands of autonomous agents opening PRs continuously and you get runner saturation, cold Docker builds on every branch, cache thrash, and a merge queue that starts behaving like a serialized database lock where time-to-commit becomes the actual bottleneck. Madison Faulkner and Hugo Santos (Namespace) lay out what replaces it: no PRs, just intent and plan fed into an agent loop with fast inline validation. Changes queue in a premerge layer where humans review intent-plus-outcome rather than diffs. The end state they're pointing toward is agents exploring multiple commits in parallel for the same plan, a multiverse where the tip of the repo is a moving target and the inner loop needs to be stateful and fast enough to keep up. Speaker info: - https://x.com/madsfaulkner - https://www.linkedin.com/in/madisonhfaulkner/ - https://namespace.so/blog/introducing-namespace - https://www.linkedin.com/in/hugomgsantos/ Timestamps 0:07 Introduction and speaker bios 1:28 Why agentic software is breaking traditional CI/CD 1:59 The fragmented lifecycle of modern software development 2:25 How traditional CI/CD pipelines work 2:55 The problems with CI/CD at agent scale 4:04 Replacing CI/CD with acceleration and orchestration 6:12 Real-world solutions and the future of agentic loops 7:23 The role of the human as the agent 8:43 Why Pull Requests (PRs) are becoming a bottleneck 10:00 A new architecture: Intent and plan-based development 11:58 Moving toward fully automated internal/external validation 13:46 The premerge queue and human-in-the-loop review 15:20 The future: Parallel development in the multiverse 16:51 Conclusion: The shifting role of CI and governance