CI/CD Bottlenecks Explode at Agent Scale
Traditional CI/CD handles humans submitting 1-2 diffs weekly with warm local caches and predictable PR reviews, builds, tests, and merges. Agents shatter this: they flood systems with thousands of short-lived branches across n repos, causing runner saturation, cold Docker builds on every branch, cache thrash from pulling the same codebase in conflicting directions, and merge queues acting like serialized database locks where time-to-commit dominates. GitHub activity confirms—commits and lines added/deleted spiked massively in recent months. Human reviewers drown (already overwhelmed), and merging divergent versions becomes impossible, turning opportunity-to-merge into the core slowdown as change rates explode.
PRs as units of work expect delayed human feedback and discrete handoffs, but agents demand inner-loop validation: regressions, compilation from known sources, conflict detection. At human scale, merge latency hides behind low volume; machines amplify it since lock times stay short but queues balloon.
Agent Loops Replace PRs with Inline Validation
Ditch PRs for intent-and-plan specs (codified in Linear tickets or Slack) fed into agent harnesses like AMP, Cursor, or Factory. The loop: agent checks out a well-known commit, implements incrementally using repo assets, runs internal validation (build/test), reports to human for 'continue' or tweaks—looping rapidly. Finished changes hit a fast merge queue into the git ledger (repo tip).
Teams at forefront (Namespace, FAL, Zed, RAMP) already operate this way, boosting volume 4x over past PR levels—humans can't review individually. Humans were the original agents: intent → loop (fix PR format, failed tests, reviewer API feedback, merge queue conflicts) → merge. Accelerate to machines, and validation shifts inside the loop for continuous enforcement of invariants like compliant checkouts.
Continuous Compute Enables Stateful, Parallel Exploration
Near-term (weeks/months): code generation speeds up (inference accelerates), internal validation must hit sub-minute (no 15-45min tests delaying loops), external validation by specialist agent LLMs (security, API conformance) providing feedback in-loop—no humans. Requires stateful environments: memory/state for incremental work, avoiding scratch restarts that kill speed. World signals (plan changes, concurrent merges) trigger plan adaptation and new loops.
High-volume changes queue in premerge layer: reconciles parallel agents on same codebase for serializability, grouping semantically (e.g., multiple commits into one reviewable bundle). Humans approve intent + outcome (feature video, LLM security outputs)—not diffs.
Future multiverse: agents tackle same plan across multiple commits (repo tip moves too fast), exploring candidates in parallel—demands ultra-fast inner loops, exploding resource use. Efficiency via incrementalism (like local workstations), warm caches as orchestration, ingress shaping/rate-limiting, agentic identity/retries. CI principles persist (validation invariants) but embed in loops; coordination lifts to harnesses coercing compliance; governance codifies in processes.
Acceleration starts over GitHub Actions: hardware/software co-design for cache-orchestrated routing. Experts like Mitchell Hashimoto (ex-HashiCorp) echo—evolve GitHub for cloud/inference scale, prioritize AI/agent users with friendly code storage.