← → navigate · S speaker tags · N notes · F fullscreen
Open Mercato
// AI Engineering

Fire and Forget

Enterprise software with zero human coding

Stan Świątkiewicz Developer Relations · Open Mercato
Patryk Tomczyk Software Architect · Open Mercato
React India 2026 · Goa · Main Stage
Stan + Patryk
01 — The problem

The promise:
agents that never sleep.

Your timeline · every day
1,000 agents running in parallel. 24/7. I haven't written a line of code in months. 🚀
♡ 12.4K↻ 3.1K💬 892
Stan
01 — The problem

The reality:
you're the babysitter.

Waiting for approvalsession 3 · 12 min
Asked a questionsession 7 · 26 min
Tests red, retryingsession 1 · 4th try
Idle. Done?session 12 · no idea
Needs a loginsession 5
Stuck in a loopsession 9 · 40 min
Rewrote the wrong filesession 2
Context fullsession 11 · restart?

Somebody has to jump from tab to tab and keep each one on track.

Stan
01 — The problem

Every session drifts
its own way.

one task"Add invoices to the CRM"
→
agent 1adds a new date library
agent 2renames the API
agent 3invents a folder structure
agent 4"refactors" auth on the way
→
merge dayFour PRs that don't fit together.
Stan
01 — The problem

More output.
Fried brain. Big bill.

brain ↓context switching all day, burned out by Thursday
PRs ↑more than anyone can review
$ ↑tokens for every retry and every lost context
Stan
01 — The problem

Companies want it too.

today Millions a year on SaaS they can't bend.
CRMERPWMSticketing+ 40 more
what they ask us "Can't AI just build it for us? Like those agents on X?"
Stan
01 — The problem

The demo works.
Then real life starts.

greenfield A new app in an afternoon. Empty repo, no rules, nothing to break. Looks amazing.
brownfield Your existing codebase. Or the demo, six months later. Conventions, history, other teams. That's where it falls apart.
Stan
01 — Today

What you'll take home.

01Agents 24/7 Fire and forget. No babysitting.01
02Prototype → production A system that still works when it grows.02
03Brownfield Your existing project, in any language.03
Stan
02 — How we got here

We had this
at scale.

140+
developers, juniors to architects
300
PRs on our busiest day

Hundreds of agent sessions.
One codebase.

Pull requests list
our pull request queue
Stan
02 — How we got here

Two rules we learned
the hard way.

rule 1 Decide before the code. A spec first. Otherwise every agent guesses what you meant, differently.
rule 2 Agents check how. Humans decide what. Is the code good? Agents. Should it exist at all? A human.
Stan
03 — Fire and forget, for realCezar

One task.
Day and night.

the task
"We have a lot of unpaid invoices. Find out why."
day · you decidefind the problem, write the issue, approve the spec
night · agents buildbuild, check, review, QA
Cezar Cezar open-source agent manager
  • starts every step for you
  • notices when one stops
  • runs all night, any agent
Patryk
03 — Fire and forget, for realCezar

One task.
Day and night.

the task
"We have a lot of unpaid invoices. Find out why."
day · you decidefind the problem, write the issue, approve the spec
night · agents buildbuild, check, review, QA
Cezar Cezar open-source agent manager
  • starts every step for you
  • notices when one stops
  • runs all night, any agent
Patryk
03 — The token question

"All these steps? How many tokens is that?"

Code is cheap.
Building the wrong thing isn't.

define the problemKnow why before you build
decide what you wantWhat's in, what's out, before any code
right-size every stepBig model to think, small ones to build. Checks cost zero tokens.
Patryk
03 — During the day · dispatch
dispatchissuespec18:00buildcheckreviewQAnight08:00

"Find out why."
Cheap agents search. The big one judges.

big model · plans
"We have a lot of unpaid invoices. Customers pay late or not at all. Find out why." → 4 suspects
→dispatch
small models · search ×4
invoicessearching…
CRMsearching…
repliessearching…
reminderssearching…
→findings
big model · judges
verdict Where is the problem? → your guess first
Patryk
03 — During the day · dispatch
dispatchissuespec18:00buildcheckreviewQAnight08:00

"Find out why."
Cheap agents search. The big one judges.

big model · plans
"We have a lot of unpaid invoices. Customers pay late or not at all. Find out why." → 4 suspects
→dispatch
small models · search ×4
invoicespaid late, not never
CRMsales chase by hand
replies"sorry, I forgot"
remindersnone sent
→findings
big model · judges
verdict Customers forget. → Remind them. → becomes an issue
Patryk
03 — During the day · issue
dispatchissuespec18:00buildcheckreviewQAnight08:00

From finding
to issue.

found problem "Customers forget to pay."
→
/om-prepare-issue One issue, with acceptance criteria and labels. No duplicates.
→
● open #6120 Remind customers about overdue invoices
featurepriority-medium
✓ why ✓ what how · next
Patryk
03 — During the day · spec
dispatchissuespec18:00buildcheckreviewQAnight08:00

Decide
before the code.

/om-auto-write-specbig model: writes the spec as its own PR, where you decide the implementation details
/om-pre-implement-specaudits it: go or no-go, before any code
you comment · you approve · you merge
Spec: invoice payment reminders #6128
Merged spec only · no code
🤖
om-auto-write-spec commented
Open questions. I picked defaults:
  • channel: email
  • when: 3 days after the due date
  • stop: once it's paid
Reply to change any of them.
👥
The audience commented
Make it 5 days.
3 days
5 days
🤖
om-auto-write-spec commented
Spec updated: 5 days after the due date.
✓Patryk approved these changes
↳Patryk merged the spec
Patryk
03 — One night · 18:00
dispatchissuespec18:00buildcheckreviewQAnight08:00

Nobody pressed start.

automation = trigger → workflow
whenAn issue gets the label agent
→
runship-it
whenA new pull request is opened
→
runreview-pr
NO FORM NEEDED
"When an issue gets the label agent, run the ship-it workflow."

GitHub · Jira · Linear · or a schedule

Patryk
03 — One night · 18:00
dispatchissuespec18:00buildcheckreviewQAnight08:00

This file
is the process.

  • each step is a skill: a Markdown file
  • each step picks its own agent and model
  • a check decides when it's done
  • same run, every time, for everyone
name: ship-it
steps:
  - id: build
    skill: om-auto-implement-spec
    runner: codex
    model: luna                 # do
  - id: check
    command: "pnpm build && pnpm test"
    onFail: { retry: build }
  - id: review
    skill: om-auto-review-pr    # fixes until it passes
    runner: claude
    model: opus                 # think
  - id: qa
    skill: om-auto-qa-pr        # fails → flagged for you
Patryk
03 — One night · build
ship-it.yaml▸ build:om-auto-implement-spec# lunacheck:pnpm build && pnpm testonFail: { retry: build }# red → build againreview:om-auto-review-pr# opusqa:om-auto-qa-pr# browser

Peek in,
any time.

/om-auto-implement-specplans from the spec, builds phase by phase

Once it starts, it asks no questions. It won't stop waiting for you.

  • every step, tool call and test run
  • tokens and cost, live
  • send a note, it reads it next turn
Cezar live run
Cezar · a live run: steps, tool calls, tests, tokens, cost
Patryk
03 — One night · check
ship-it.yamlbuild:om-auto-implement-spec# luna▸ check:pnpm build && pnpm testonFail: { retry: build }# red → build againreview:om-auto-review-pr# opusqa:om-auto-qa-pr# browser

Tests fail.
Nobody gets paged.

check   pnpm test        ✗ 2 failed
        reminder.test.ts › skips paid invoices
build   retry, with the error
check   pnpm test        ✓ 12,214 passed
check   e2e, fresh app + db ✓ green
the rule The tests say done.
Not the agent.
Patryk
03 — One night · review
ship-it.yamlbuild:om-auto-implement-spec# lunacheck:pnpm build && pnpm testonFail: { retry: build }# red → build again▸ review:om-auto-review-pr# opusqa:om-auto-qa-pr# browser

The author never
reviews its own code.

/om-auto-review-prreviews against your CODE_REVIEW.md, fixes until merge-ready
different agentFresh session, fresh worktree
your checklistCorrectness, security, breaking changes
request changes?Fix, test, review again

build: cheap model · judge: big model

Patryk
03 — One night · QA
ship-it.yamlbuild:om-auto-implement-spec# lunacheck:pnpm build && pnpm testonFail: { retry: build }# red → build againreview:om-auto-review-pr# opus▸ qa:om-auto-qa-pr# browser

It clicks through
the app.

/om-auto-qa-prwalks the change in a browser, posts a report
  • real browser, fresh app and database
  • happy path and the edge cases
  • one comment on the PR, screenshots attached
om-auto-qa-pr commented on #6131QA passed
overdue 5 days✓ reminder sent
invoice already paid✓ nothing sent
no email on file✓ skipped, logged
console✓ no errors
From: BillingInvoice INV-2041 is 5 days overdue

Hi Ana, a friendly reminder: invoice INV-2041 (€1,240.00) was due on 19 Oct.

Pay now
Patryk
03 — One night · night
dispatchissuespec18:00buildcheckreviewQAnight08:00

You sleep.
The manager doesn't.

  • many workflows at once, one per slot
  • a slot is free? the next workflow starts
Cezar on a phone: tasks
Cezar on a phone: session
Patryk
03 — One night · 08:00
dispatchissuespec18:00buildcheckreviewQAnight08:00

Good morning.
You press merge.

/om-merge-buddywhat can merge now, what's blocked, and why
nightQA left a summary and screenshots on each PR
08:00Coffee. The merge skill lists what's ready: #6120 · #6124 · #6127
readOne or two comments
mergeClick merge.
the one hard rule No human approval, no merge. QA failed? It's flagged on the PR, not hidden. You decide.
● Ready to merge · waiting for a human0 humans approved✓ Merged
Patryk
03 — The glue

Many agents.
No collisions.

own worktreeNobody touches anybody's files
claim firstAn in-progress label. Others back off.
on GitHub · GitLabEvery step visible where you already work
Patryk
03 — Wrap-up

Fixed process.
Swappable models.

day · you decide → night · agents build → morning · you merge

Same steps and checks for every task. Each step gets the model it needs.

Patryk
03 — Under the hood · models today, October 2026

Our picks today.
Not forever.

Job Today's pick
Heavy thinking: specs, judging Opus 5 · Fable (Anthropic)
Cheap execution: build, search Luna (OpenAI): clearly beats Haiku
Security fixes Codex · Sol or Astra
Sensitive data, nothing leaves the building self-hosted models
Zero budget free models via OpenCode
Stan
03 — Under the hood · model-agnostic

In six months?
Who knows.

Change one line. Keep the tool, the skills, the workflows.

Claude CodeCodexCopilotCursorJunieOpenCodePi
Stan
04 — Available now · fully open source
Cezar
npx cezar-run

or try the desktop app

Skills built in.
Cezar task list
Cezar · parallel tasks, each in its own worktree
Stan
04 — Newest update

Fresh in Cezar.

newWorkflows
newAutomations
new · for power users Unlimited accounts Many subscriptions of the same provider, all at once. A default per project.
Claude
Company · Maxdefault
Personal · Pro
Client project · Max
Codex
Company · Business
Personal · Plus
OpenCode
Free models · no plan
Stan
04 — Brownfield

Old code base?
Doesn't matter.

.NETJavaPHPPythonGoyour 10-year-old monolith
bring your processYour labels, your checks, your rulesOverride one skill. Keep the rest.
we'll helpSetting it up at your company?Talk to us.
Stan
Thank you · it's all freeCezar
Patryk Tomczyk
Patryk Tomczyk
Software Architect
in/patryk-tomczyk github.com/patzick

It's all free
and open source.

QR: github.com/open-mercato/cezar
cezar.run · github.com/open-mercato/cezar
★ 811 · give us a star ⭐
Stan Świątkiewicz
Stan Świątkiewicz
Developer Relations
in/stan-swiatkiewicz
Stan