We took the manual overhead out of engineering.
AI agents now do the slow, repetitive work that used to cap what we could take on — so we ship more, faster, with proof it's correct, including work we used to simply defer.
Manual → prompt → agent → multi-agent → overnight.
A prompt becomes an agent; one agent becomes a fleet; the fleet runs itself overnight. Each leap moves people further from doing the work and closer to just deciding — here’s what changes.
You do the whole job by hand.
You write the instructions once; the AI does the task and hands it back.
The prompt becomes a dashboard that runs the whole job and checks itself.
Many dashboards — a specialized agent for every recurring job (30+).
The agents run on their own, unattended, and report back by morning.
By hand.
You open the change and read every file, line by line, until you spot the problem — then you write the comment. Thorough, but it’s hours, on every single change.
So we wrote it down — a “prompt.”
Instead of reading it all yourself, you write the instructions once. The AI reads the whole change and writes the review back. You read it and decide — minutes.
- bug
getUsercan dereferenceundefinedwhen the id isn’t found (line 14). - style use strict equality
===on line 13. - nit
listUsershas no guard for a null filter.
Then — one click.
We wrapped that prompt in a dashboard. Paste the change, press one button: it deploys a live preview, takes before/after screenshots, and writes the whole review. Seconds of your time.
Then — one for every job.
That one dashboard worked so well we built one for each recurring task. Today it’s a fleet of 30+ specialized tools — each still just seconds of your time, but now covering every job we do again and again.
Now it runs itself.
No one clicks anything. Overnight, agents pick up every open change, review them all, and the board fills itself in. By morning it’s done — zero effort.
AI changes what we can say yes to.
Sort any task by how hard it is and how long it takes. Before AI, three of these four boxes got deferred. Now agents do the work and people do the review.
The work we do — and who does it now.
These are jobs we used to do entirely by hand. Sort them by how hard and how slow they are — then hand them over and click any one to see the tool that now runs it.
capture in pass 2
The results, in numbers.
Agents raise the dependency bumps, validate them in a live preview, and ship the fix — clearing the CVE backlog.
▼ 94% of known CVEs clearedCustomer (JIRA) and small (mixin) issues assigned to me and worked by agents — 17 JIRA + 7 mixin — that would otherwise sit unhandled.
▲ 9 fixed, the rest in progress · click for the breakdownHardened across one batch (a second followed) — the failure rate fell to near zero, now runnable unattended overnight.
▲ ~99% attempt-1 pass rate
One control plane manages them all — purpose-built AI consoles, one per recurring job.
▲ and growingA weekly agent ages out issues that have gone quiet — 129 flagged, 101 closed — so the backlog reflects real, active work with no manual sweep.
▲ hands-free backlog hygiene · click for detailEvery new issue is read, questioned for missing detail, and flagged ready for triage — humans open a triage-ready queue instead of raw reports.
▲ triage-ready, hands-free · click for detailAgents write the missing unit tests every day — steadily lifting coverage on a codebase that had gone years under-tested.
▲ +7.4 pts coverage · click for the graph“But how do you know it’s right?”
We don’t take the AI’s word for it. Every change is checked the way a person would check it — and the AI shows its work.
- 01
It drives a real browser
The agent clicks through the actual product, like a user would.
- 02
It captures the evidence
Before/after screenshots and a screen recording of every run.
- 03
It checks against a rulebook
Automated tests and tagged, labelled checks confirm the expected result.
- 04
A person signs off
Nothing ships on trust — a human reviews the evidence and approves.
Built to be safe and to last.
This isn’t ad-hoc. We treat the AI system like any other piece of infrastructure — with security, standards and a long-term plan.
Sandboxed by default
Agents run inside isolated containers — no direct access to private or protected company data.
Shared best practices
A common library of prompts and rules so everyone works to the same standard.
Architecture mindset
We weigh long-term trade-offs, not just quick wins — the same way we build products.
A dev dashboard future
Consolidating the tools into one place: our long-term investment, not a fad.
10 minutes of it actually working.
The most convincing part isn’t a slide — it’s watching an agent do the job. Here’s the running order (with a recorded version as backup, in case the network isn’t kind).
- AAn agent reviews a code changeDeploys a preview, screenshots before/after, writes the review — live.
- BAn agent clears a vulnerabilityFinds it, fixes it, proves the app still works.
- CThe dashboard fleetA quick tour of the tools we run every day.