Your CI starts at the PR.
zeroshot starts at the issue.
A meta-harness for coding agents, built for teams that write code faster than they can review it.
Open the platform try for freeBook a call with a founderGitHub
npm install -g @the-open-engine-company/zeroshot
- Microsoft
- Amazon
- NVIDIA
- Meta
today with zeroshot
ana claude code + her prompt ──┐
ben codex + skills_v3_final ───┤
kai copilot + "be careful" ────┤ ana ben kai mo
mo pi + whatever works ───────┤ └────┴──┬─┴────┘
│ │
┌──────┴──────┐ ┌──────┴──────┐
│ 47 open PRs │ │ zeroshot │
└──────┬──────┘ └──────┬──────┘
│ │
you, 1:40am, LGTM × 47 every change tested + reviewed
│ │
"tests pass," says the agent "tests pass," say your checks
│ │
prod, PagerDuty at 3:12am prod, quiet at 3:12am
token bill ↑↑↑, owner unknown you, asleep. one process, whole team
today
ana claude code + her prompt ──┐
ben codex + skills_v3_final ───┤
kai copilot + "be careful" ────┤
mo pi + whatever works ───────┤
│
┌──────┴──────┐
│ 47 open PRs │
└──────┬──────┘
│
you, 1:40am, LGTM × 47
│
"tests pass," says the agent
│
prod, PagerDuty at 3:12am
token bill ↑↑↑, owner unknown
with zeroshot
ana ben kai mo
└────┴──┬─┴────┘
│
┌──────┴──────┐
│ zeroshot │
└──────┬──────┘
│
every change tested + reviewed
│
"tests pass," say your checks
│
prod, quiet at 3:12am
you, asleep. one process, whole team
Turn your best engineer’s judgment into a process.
Enforce your own QA workflow as a cyclic graph
We wrote a blog post about how the SDLC changes when agents build, test and review the code.
imagine shipping as fast as you can plan.
illustrative run · queue issues with dependencies; zeroshot runs every one through your QA workflow
go to bed :)
today
GitHub
47 pull requests are waiting for your review
Slack · ben in #ai-tools
can you send me the skill?
Slack · kai in #eng
you have to run the /test-performance command
Slack · mo in #ai-tools
why are you using fable for this?
Slack · ana in #reviews
LGTM pls, it's only 4,000 lines
Slack · kai in #ai-tools
there are 3 CLAUDE.md files and they disagree
Slack DM · ben
can you just approve my 14 PRs? standup is in 6h
Slack · ana in #eng
i have to go home now. i'll review it tomorrow
Email · finance
why is the AI bill bigger than payroll?
Slack · ben in #ai-tools
skills_v4_final_FINAL or skills_v4_final_actually?
GitHub · mo
merged #90: “tests pass on my machine”
Slack · kai in #eng
who the hell rewrote the system prompt?
Email · alerts
Checkout error rate: 94%
Slack · ben in #incidents
wtf, who merged #90??
Slack · mo in #incidents
i told my fucking agent to test it omg
Slack DM · ana
classic mo
Email · support
214 tickets: customers can't pay
Email · legal
a customer just said the word lawsuit
X
@yourapp down again?? switching vendors
Slack · kai in #incidents
fuck. checkout is down
PagerDuty
Critical: checkout is down
Slack · ana in #incidents
@channel is anyone awake??
Slack · ben in #incidents
who the fuck approved #90
Slack · mo in #incidents
it worked on staging tho
PagerDuty
Escalated: still not acknowledged
Slack · kai in #incidents
shit. the rollback failed too
Phone
PagerDuty is calling you
Email · ceo
call me. now.
Email · investor
saw the outage on X. quick call?
Calendar
9:00 AM · Postmortem: checkout outage
a recruiter: are you open to new roles?
0 PRs shipped1 PR shippedit broke checkout.
with zeroshot
0 PRs shipped1 PR shipped2 PRs shipped3 PRs shipped4 PRs shipped5 PRs shipped6 PRs shipped7 PRs shipped8 PRs shipped9 PRs shipped10 PRs shipped11 PRs shipped12 PRs shipped13 PRs shipped14 PRs shipped15 PRs shipped16 PRs shipped17 PRs shipped18 PRs shipped19 PRs shipped20 PRs shipped21 PRs shipped22 PRs shipped23 PRs shipped24 PRs shipped25 PRs shipped26 PRs shipped27 PRs shipped28 PRs shipped29 PRs shipped30 PRs shipped#90 too. tests caught it first.
see it work.
Claude Code on Opus 5.5 builds the page from the same README in all three.
one session, no reviewer
Its final report
“I checked it with headless-Chrome screenshots at phone, tablet and desktop widths, in light and dark mode, and with reduced motion.”
one session, no reviewer
Its final report
“Checked in the rendered page: desktop, tablet and phone layouts in light and dark, with no horizontal scroll from 320 to 1920 px.”
What it shipped
- Eight screens, whereas the reviewed page is five.
worker → design slop reviewer → repair
a custom graph for this experiment
Round 1: sent back
- Too long: the README poured into ten screens.
- Apart from the demo, it could be any dev tool’s page.
- On phones the staged demo run loses its “illustrative” label and passes as real.
Round 2: accepted
Five screens instead of ten.
An arctic flight game, built with and without zeroshot.
With zeroshot
The zeroshot graph used for this experiment
Run start
Candidate 1
Candidate 2
Candidate 3not just
for coding.
Autoresearch runs experiments through three independent judges.
Explore autoresearch
Watch a research run · 20 sec
what you get.
- agents
- Claude Code · Codex · Copilot CLI · Pi, mixed per step. On your own keys, direct or via OpenRouter, Bedrock or your own gateway.
- ships to
- GitHub · GitHub Enterprise · GitLab
- runs on
- Our cloud: every run gets its own environment, in parallel, off your engineers’ laptops. AWS in Frankfurt. Or your own machine, on your Claude, ChatGPT or Copilot subscription: the engine is open source.
- security
- SSO · Org roles · DPA · we don’t train models on your code

