Team of one

I have zero direct reports.
My team is 8 AI agents.

I'm a feature owner on a consumer mobile app. Over four months I built a crew of agents that run the daily analytics, the prototypes, the bug triage, the meeting notes and the small code fixes. This page has no company metrics in it — only hours, and how much faster things got.

8 agents, all mine 13 scheduled automations Claude Code + MCP human in the loop
My own week, back
27h / week
Roughly two thirds of a work week that used to go into SQL, dashboards, meeting notes, ticket writing and triage.
Conservative estimates of my own recurring work.
Other teams' hours are counted separately below.
8
agents I wrote myself
13
automations on a schedule
~45 sec
to file a ready ticket
57 min
bug report → signed build
0
developer hours on that fix
24/7
triage bot, off my laptop
01 — the ledger

Where the hours actually come from

Every row is recurring work I used to do by hand every week. "Now" is what's left for me: reading, checking, deciding.

Recurring workWasNowSaved / week
Daily analytics & anomaly watch ~6 h of SQL and cross-checks ~1 h — I read the alerts ~5 h
Dashboards for demos & sprint reviews ~4 h of pulling exports 0 — they refresh themselves ~4 h
Walking the app, mapping the journey ~3 h clicking through screens ~0.5 h — I read the report ~2.5 h
Meeting transcripts → action plans ~4 h of note-taking ~0.5 h — I approve the plan ~3.5 h
Reporting: standup, sheets, weekly demo ~3.5 h of filling cells ~0.5 h — I sanity-check ~3 h
Writing tickets in the tracker 20–30 min × ~19 tickets ≈ 8 h ~0.25 h — 45 sec per ticket ~7 h
Bug triage out of chat ~1.5 h of "which version? got a screenshot?" 0 — the bot does the interview ~1.5 h
My time, per week ≈ two thirds of a work week ~27 h
The point isn't the total. It's what the freed hours turned into: more experiments in flight, more prototypes looked at on a real device, more decisions made on data instead of vibes.
Engineering time
Small bug fix: 4–6 h of a developer + days in the backlog0 dev hours, a build the same day.
Design time
Feature prototype: 2–5 days in a designer's queuea clickable page in hours, on my phone.
Data team time
Ad-hoc data question: days waiting in a queueminutes, straight out of the warehouse.
02 — speed

Saved hours are nice. Shorter loops change the job.

When a thing takes minutes instead of days, you stop rationing it. That's the real shift.

Filing a ticket
20–30 min~45 seconds
I say what's broken out loud; the agent writes it up in my team's template — sprint, epic, labels, acceptance criteria, screenshots inline, both languages. I review and hit send.
Asking the data a question
days in a queueminutes, from my phone
The analytics agent has a Telegram bot. I type "why did conversion dip yesterday?" while walking, and get an answer computed off production events — with the de-dup and anti-fraud rules already baked in.
Seeing a prototype
2–5 dayshours
The design-engineer agent ships live HTML/CSS prototypes and deploys them, so I open the real thing on a real phone instead of arguing about a static mockup.
Nothing waits for me
"I'll get to it"13 scheduled runs
Anomaly watch, dashboard refreshes, meeting plans, the evening standup post — done on a timer before I open the laptop. The bot that triages bugs runs on a server, not on my machine.
One bug, start to finish
57 minutes · 0 developer hours
  1. 10:35 Agent writes the ticket: current → expected → tech notes → acceptance criteria, screenshots inline
  2. 10:36 Comment /ai run kicks off the coding agent in the real repo
  3. 10:40 Pipeline running against the production Android codebase
  4. 11:05 Agent reports back: cause found, one-line fix, self-review, risk notes
  5. 11:07 Comment /build
  6. 11:32 Signed APK posted on the ticket. QA confirms it by 14:05
To be precise about credit: the coding pipeline is my team's, built by our platform engineer. What's mine is the agent that drives it end to end — writes the spec, launches the run, watches the PR, orders the build, then pings me with the link. I've taken 26 tickets through it that way. None of them cost a developer an hour.
03 — the crew

Eight agents, one job each

Not one mega-prompt. Specialists with their own tools, their own access and their own schedule — the way you'd staff a small product pod.

B
Boris
Head of analytics
Single front door for anything data. Routes the question, and for "why did it drop and where do we lose people" runs both analysts and merges the answer.
on request
A
Anatoliy
Metrics analyst
Queries the warehouse, hunts anomalies daily, keeps the dashboards and the reporting sheets current. Answers questions in Telegram.
daily 12:00 + 24/7 via chat
A
Anton
Journey analyst
Walks the app on a real phone, screenshots every screen, builds the funnel from actual steps, compares onboarding versions, benchmarks competitors.
on request + playtests
P
Petya
Design engineer
Live HTML/CSS/JS prototypes, onboarding screens, landing pages, redesigns — deployed to a URL so I can check them on a device in minutes.
on request
C
The Council
Hypothesis factory
Five different frontier models argue with each other for several rounds — grounded in our real numbers and real screens — then a red-team pass, then a synthesis with priorities.
on request
R
Rigger
Pipeline operator
Takes a bug end to end: writes the ticket, launches the coding agent, follows the PR, orders the build, pings me when there's something to install.
runs the whole lane
S
Secretary
Meeting assistant
Turns any call transcript into decisions, owners and deadlines, and drops the plan in chat. Nothing agreed in a meeting quietly evaporates.
weekdays, every 30 min
F
Foreman
Focus enforcer
Compares what I'm actually doing against the priorities that matter. If I've drifted, it tells me so, bluntly, in Telegram. Yes — I let an AI police my focus.
hourly, 10:00–20:00
H
Helper
Bug triage bot · not counted in the eight
Anyone on the team throws a bug into chat. The bot interviews them like a QA engineer, collects screenshots, files a complete ticket and suggests an owner. Runs 24/7 on its own server, so it doesn't care whether my laptop is open.
24/7 · systemd, auto-restart
04 — how it's wired

Boring plumbing, on purpose

  • Subagents, not a mega-prompt. Each one has a narrow brief, its own tool set and its own permissions. Small scope is what makes them reliable.
  • MCP for the unglamorous part. Warehouse, product analytics, tracker, wiki, chat, design files, and a real phone over ADB. A chat window can't reach any of that.
  • Schedules beat good intentions. Thirteen jobs on timers. The morning work is finished before I sit down.
  • Method lives in a repo, not in my head. One source of truth for how metrics are defined — cohorts, de-dup, fraud filters — so every agent counts the same way I would.
  • Packaged for other people. Two of the agents are already stripped of my credentials and handed to colleagues with a setup wizard. An agent only one person can run isn't leverage, it's a hobby.
Claude CodesubagentsMCP skillsscheduled tasksClickHouse feature flags + A/BAndroid over ADBCloudflare Pages
05 — what stays manual

Roughly a quarter of my week is still by hand

  • Product decisions and prioritisation. That's the job. Delegating it would be delegating accountability.
  • Alignment with people. Stakeholders, live meetings, disagreements. Nobody wants a negotiation with my agent.
  • Review on anything that matters. Money numbers, production code, messages going out to humans — a person reads it first. Every time.
  • Final acceptance on a real device. A build isn't done because an agent says it is.
This isn't a backlog of things I haven't automated yet. I automate execution and keep the judgement. That split is the whole design.

Automation isn't a pile of scripts. It's an operating model: one person decides, agents do the rest.

Konstantin Kutovoy · feature owner, consumer mobile app · 2026
All figures here are hours and cycle times from my own week — conservative, self-measured, and rounded. No employer metrics, revenue, user counts or product names appear on this page by design.