Fleet: why every NKE engineer ships with a crew
Fleet: why every NKE engineer ships with a crew
Most staffing pitches stop at the person. Here's the résumé, here's the rate, good luck.
Ours doesn't, and this post is about why. Every engineer NKE places comes with Fleet — the agent harness we built for running AI coding workers on real engineering tasks. It's the reason a single senior engineer from our bench can carry the throughput of a small pod. I want to explain what it actually is, because "our developers use AI" is the emptiest sentence in the industry right now, and I'd rather show the machinery.
What Fleet is
Fleet is a harness for running headless AI coding agents — Claude Code workers, in our shop — under real engineering discipline.
The flow looks like this. An NKE engineer scopes a task the way they'd brief a mid-level teammate: context, constraints, acceptance criteria. Fleet hands it to a worker that builds in an isolated branch of its own worktree, runs the repo's own linting, type checks, and tests, and opens a pull request. The engineer stays where seniors belong: on the design going in and the review coming out.
One task, one run, one PR. From the outside — your repo, your CI, your review process — Fleet's work looks like any other engineer's work. Because it has to pass the same gates.
The part nobody talks about: verification
Here's what we learned the hard way: agents lie. Not maliciously — optimistically. Early on, workers would report "done" with uncommitted changes and failing tests sitting in the tree.
So Fleet doesn't take the agent's word for anything. After every run, the harness independently verifies the actual state of the work: is the tree clean, do commits exist, are they pushed, do the repo's own quality gates pass. A run is "succeeded" only when everything is green — otherwise it's flagged incomplete, and a human looks. The agent's self-assessment is treated as marketing.
That one design decision is most of the difference between agentic coding that compounds and agentic coding that generates plausible-looking debt.
Boring infrastructure, on purpose
Fleet workers run as supervised system services on dedicated hardware — not as a script in someone's terminal. That means they survive dropped connections, they run under kernel-enforced memory limits, and when one dies, we get a postmortem instead of a mystery. We tried the lighter-weight approaches first. Dead workers with half-finished branches taught us to stop.
It's also platform-agnostic — the harness doesn't care whose model is under it — though we've standardized on Claude Code and every NKE engineer gets an account we provide.
What this means if you're hiring our engineers
A few things, plainly.
Throughput. An engineer who can delegate well-scoped tasks to verified workers ships more than one who types everything. Not 10x — anyone quoting 10x is selling something — but meaningfully more, on the boring-but-necessary work especially.
Discipline included. Agentic coding without strict linting, tests, and human review produces what I call vibe code: plausible, confident, and wrong in ways you find in production. Fleet exists specifically to prevent that. If your repo's gates are weak, we'll tell you — that conversation is part of the value.
You're in control. Fleet is opt-in per engagement. We ask before turning it on, and we turn it off if your policies say no. Your engineer is fully productive either way — Fleet is a multiplier, not a crutch.
No lock-in. Everything Fleet produces is ordinary git history and PRs in your repo, reviewed by a human you know. If we walked away tomorrow, you'd keep normal code.
What Fleet doesn't do
It doesn't architect systems, own ambiguity, or talk to your stakeholders — that's the engineer's job, and it's why we place senior people rather than selling you the tooling alone. A harness amplifies judgment. It doesn't replace it.
We run NKE on this same principle — our marketing, our publishing, our follow-ups are agents we built, supervised by humans. This blog post reached you through that pipeline. Last week Fleet ran 731 tasks and delivered 98 merged pull requests across 12 production repos. 261 runs were flagged by verification and never reached a human reviewer's desk — which is the system working, not failing.
If you're curious what a senior engineer plus Fleet would do inside your codebase: our rates are published, engagements are month-to-month, and you interview the engineer before anything is signed. What's the task your team keeps not getting to?
Joe Juhnke is the founder of NKE, Inc. — senior nearshore engineers, AI-native, vetted personally. novemberkiloecho.com