AFK Development: humans at both ends, agents in the middle.

AFK is away from keyboard. Your people align on what to build and review what came back; the implementation in between runs unattended. It is how we deliver, and it is the pipeline we set up inside client teams.

On a client engagement running this pipeline, development throughput went up 70%, and 85% of that project’s code was agent-generated. Every pull request still met a human. Read the case study.

The pipeline

Every constraint here exists because of a failure it prevents. That is the whole design.

  1. 01Human

    Spec

    Agreement on what to build, before an agent touches anything.

    Without it: every later stage compounds the mistake.

  2. 02Agent

    Vertical slices

    Work cut into thin end-to-end slices that do not block each other.

    Without it: runs drift outside the range the model is reliable in.

  3. 03Agent

    Feedback loop

    The runner runs the checks (lint, types, tests, browser E2E) and feeds every failure back until they pass.

    Without it: the signal that the code works arrives late: in CI, or from a tester days later.

  4. 04Agent

    Refactor pass

    Cleanup gets a phase of its own rather than being assumed.

    Without it: models cut corners and the codebase erodes quietly.

  5. 05Agent

    Agentic QA

    The user journey driven end to end in a real browser.

    Without it: unit tests miss what the user actually meets.

  6. 06Human

    Review

    Developer review and business sign-off.

    Without it: nothing catches what judgment catches.

Why ours runs in the cloud

The usual setup runs on a developer’s laptop, which caps you at the parallel work one machine can hold and kills a run when the lid closes. We run pipelines in cloud development environments instead, so slices run in parallel without fighting for one machine, and Claude, current stacks and current plugins stay the same across every environment.

The honest constraint: run as many pipelines as your review capacity can absorb, and no more. The bottleneck moves to human judgment, which is exactly where it should sit.

Tell us what your team keeps doing by hand.

Thirty minutes, no charge. You leave with an honest read on whether it is worth automating.