Blog · AI-assisted delivery
The delivery method behind our four commercial Cloudflare products — with two weeks of real numbers, not vendor slides.
"AI-assisted delivery" is the most inflated phrase in software right now, so we hold ourselves to evidence. A recent two-week window across our product fleet: 43 sessions, 201 commits, over 3,000 shell operations and roughly 407 hours of agent session time — spanning Cloudflare infrastructure repair, platform feature work, a documentation product build-out, an RBAC and tenancy rollout, and multi-repo restructuring. All of it production work on the products this site sells.
The point of those numbers is not volume. It is that the method holds at volume — across five parallel workstreams, in a real multi-repo estate, without the wheels coming off.
The method
Work is delegated as numbered epics with checkboxes, not chat prompts. The agent gets a whole phase — "implement E34.1–E34.7" — and the plan file is updated in the same commit that ships the code. Any session, human or agent, can reconstruct system state from the plans alone.
Agents run long and autonomously, but inside boundaries: git worktrees isolate their changes, staging is always explicit (never git add .), permission classifiers block unauthorised actions, and scope violations get flagged rather than absorbed.
A green build is not the finish line. On a recent access-control rollout the agent applied the production migration, then proved the fix by recovering a one-time password and completing a live sign-in. Deploys are smoke-checked on the live domain, every time.
Every session starts informed, not blank: project rules, platform standards, plans and fleet health are aggregated into a live layer the agents query over MCP. We built SourceAtlas because we needed exactly this — then made it a product.
One recent platform upgrade ran as a four-day plan: features implemented and deployed to production, five hundred real business records seeded, and the next phase queued — delivered as one coherent commercial product, not a pile of PRs. The human role is technical direction: setting the plan, correcting a premature diagnosis, catching the deploy an agent thought it had finished. Surgical intervention, not line-by-line review.
That is also what we install for clients: not "access to Claude", but the working system — plan formats, worktree guardrails, CI gates that block unverified work, and the documentation layer that gives agents memory. Details on the AI-assisted delivery service page; the fuller essay is on CIOatWork.
