Twelve months ago we were still one job at a time, with a person in every loop.
That was not caution for its own sake. We did not trust what the AI was doing yet. So we stayed hands-on: prompting, checking, merging, one feature at a time.
Before we went anywhere near a fleet, we built a proper context architecture across our projects. Whatever an agent produced had to pass our coding standards and avoid slop. It had to be dependable, reliable. and consistent. Only when we were comfortable with that did we move into fleet mode. We still check diligently that those standards are being followed.
Now we run agents across multiple worktrees, on multiple products. Same people. Different job.
The problem was not the agents
The hard part was losing the plot.
With several agents running at once, it is easy to forget what is going on and what is running where. Which branch is waiting on a person? Did this run follow the workflow we spent a year sharpening, or did it skip the bits that make review safe? Where are the screenshots? Which files changed?
One job with a person in the loop does not force those questions. A fleet does.
What Lowdown is
Lowdown is our internal tooling for this: a local board that gives us observability over agent work. It is not a product launch. It is how we keep sight of our own process.
That heavy human-in-the-loop habit forced us to write the steps down and tighten them until the output was trustworthy. The same discipline survived into fleet mode.
It watches every branch through the loop we use in production:
grill → isolate → build → evidence → test → PR → close out
For each branch you see the step it has reached, how long it has been there, whether an agent is on it, the files that changed, and the evidence that run produced. Anything that needs a human lights up magenta.
Lowdown reads git, run logs and evidence folders on your machine. It does not run the agents. It does not talk to GitHub. It does not touch your code. You work in the terminal; Lowdown watches.
Grill before code
Decisions get written down first: what we are building, for whom, and what is out of scope. Review can disagree with a sentence, not a vibe.
Evidence with every run
Every agent leaves a trail: screenshots, a short recording, and a file map. Review is grounded in what actually happened, not a trust exercise.
Stop at test
When the suite passes, the agent writes a hand-off (who to log in as, which pages to check) and stops. Magenta means ready for a person. A human tries the change. Only then does a PR open.
Then PR, then done
History kept
Finished branches stay readable: grill notes, evidence, timings. The loop is not only followed in the moment. You can see that it was followed.
Why observability?
When you have one agent, you can keep it in your head. When you have a fleet, you cannot.
Lowdown preserves that picture: what is running, where it sits in the process, what evidence it left, which files it touched. The tooling and the development workflow stay adhered to because the board shows drift the moment it happens.
Planning and review take longer than they used to. Build time is down. Overall it is a net win, but only because the loop is enforced by what you can see, and because we still check the standards ourselves.
Where we are
This is the first time we have talked publicly about AI in our development process. It felt worth saying out loud at this stage: how the job changed, how we earned the right to run a fleet, and what we built so we would not get lost in it.
Most teams are still in heavy human-in-the-loop mode, and that is fine. Autocomplete was years ago. Orchestrating a fleet only makes sense once the outputs are dependable. The lesson we keep repeating internally is simple: the model is not the bottleneck. Context, standards, workflow and observability are.
Metadeck is a software agency that also runs its own SaaS products. That dual view is how we learn what actually works for clients. Lowdown is our internal board for agent-assisted development on our own work.
Get in touch
If you want to hire a team at the cutting edge, with well-defined processes behind the work, drop a note below.
