Plusory digital teammates · Your team, plus one.
Everyone else sells you a tool with a dashboard. Petra is a colleague with a memory.
A digital teammate who takes the vigilance half of engineering — the work around the code that never got cheaper. What she learns about your systems stays learned.
- On watch 24/7
- Answers every page first
- Read-only on day one
- A fraction of the hire
For VP Eng and platform leads at 20–200-person teams. Three design-partner slots this quarter.
P2 — api-gateway p99 latency > 2.5 s (prod-eu). Policy layer 1 → Petra.
Looking. Spike started 01:54 — lines up with the nightly export job. Error rate is flat; customers unaffected.
Likely cause: the export job is holding a table lock. I changed nothing — holding for morning. Graphs and repro are on the ticket.
No human woke up.
Nightly run is in. 14 failures, sorted — three minutes to read.
| Test | Verdict | Evidence |
|---|---|---|
| checkout-smoke | Flaky | 3 reruns green |
| installer-upgrade-vN1 | Real | repro + recording |
| login-sso-regression | Recurring | 3rd since May — same root cause |
- Overnight
- 14 failures — 11 flaky (reruns green), 2 real (repros attached), 1 recurring: login-sso, third time since May, same root cause.
- Today
- Watching the 4.2 release candidate.
- Blocked
- Staging creds expire tonight — who can refresh?
Refreshing the staging creds now — new expiry lands in the vault in 10.
Watched the admin-console walkthrough. Filed 12 steps into my app notes — they stay learned.
- Refunds live under Billing → Adjustments, not Orders
- Legacy console logs out after 90 s idle — noted a retry
- Two steps need a second approver — flagged, not attempted
Her job description
Hire a duty engineer: on watch around the clock — tests every build, watches every dashboard, answers every page first, signs off every release.
Her desk is every surface you have: browser UI, desktop apps, terminals and legacy consoles, the dashboards you name, chat and the pager.
-
01
Morning triage
Mornings start with answers. Overnight failures arrive sorted — flaky vs. real vs. same-root-cause — with repro steps and recordings, before standup.
→ her standup update is in the channel before you sit down
-
02
Bug reproduction
Every escalation arrives with evidence attached: repro, logs, suspect area. A vague ticket comes back as steps someone can follow.
→ repro + recording on the ticket
-
03
24/7 watch
She takes layer 1 of your existing escalation policy; humans stay layers 2 and up. She watches the dashboards you name, the log tails and the consoles, on deterministic probe cadences — not on a hunch, and not only when someone remembers to look.
→ escalation with evidence — or quiet
-
04
On-call first response
Pages get answered in minutes. She correlates the page with what changed — the deploy, the cron, the config — and wakes a human with that already attached, or not at all.
→ diagnosis attached to the page
-
05
Release sign-off
Releases ship with a sign-off, not a shrug. An evidence-linked safe/not-safe report for every release candidate.
→ safe / not-safe, evidence-linked
How hiring works
You hire her the way you already hire a contractor. There is nothing to install and nothing to integrate.
-
She gets a workstation
A Windows VM you provision and a login you issue — the same first-day checklist you already use for a contractor.
-
Your credentials, your environment
She works inside your environment, under credentials you issue and can revoke, on the surfaces your team actually has — including the desktop tools and legacy consoles.
-
Her reports land where you already read
In your team chat and on the ticket — not in one more dashboard somebody has to remember to open.
Why “hire” is the right verb
Every product in this space is a tool: you operate it, it forgets you between sessions, and it waits to be told what to do. Three traits make her an employee instead — each one exists because it makes a duty you’re paying for better.
-
She remembers
The test that’s been flaky since spring. The installer bug back for the third time. Her month-three triage is sharper than her week-one — a session-scoped tool meets your codebase new every run.
-
She’s where your team is
Anything that can send a webhook or an email can page her — your CI, PagerDuty, Grafana alerts, a cron job. She reports back in your team chat, in writing you can forward.
-
She doesn’t wait to be told
A threshold crossed at 2 a.m. becomes an investigation, not a forwarded alert. She looks, reproduces what she can, and either handles the night quietly or wakes a human with the diagnosis attached.
Why this is safe to try
-
Earned, not assumed
Read-only on day one — capabilities are granted like any new hire’s. Every capability is yours to grant, and yours to take back.
-
The audit record
Every action is written to an audit record as she does it, built to survive a crash with the record intact. Bring your security team; we like those meetings.
-
Being wrong is cheap
Her day-one failures cost you nothing, because she’s read-only. And being wrong stays cheap after that: every conclusion links the evidence it came from, so a bad call is visible in seconds rather than discovered in production.
- write-ahead auditevery action recorded as it happens, not after
- allowlistwhat she can do is enumerated, not open-ended
- deny-listirreversible actions blocked by default, on day one
- drift detectiona changed UI makes her adapt, not get stuck — and never produce a silently wrong output
- self-healshe repairs a broken step and reports the healed knowledge for review
- crash durabilityshe recovers without double-messaging a human
- revocable grantsoffboarding is a login switched off, in a minute
Anything not written here — retention terms, log export, update windows — is agreed per customer during onboarding. We would rather say that than invent an answer to get through your review.
Talk to us about a pilot
Tell us which duty is costing your team the most. If Petra isn’t the right fit for it, we’ll say so — and point you to what is.
Jiri Mares — ex-Microsoft, ex-Amazon, 15 years building and running software.