The dark factory for code
Hyper spawns N coding agents in parallel, independently verifies every output, and merges only the winner.
$ hyper run --spec add-oauth --agents 4
Evaluating 4 candidates...
agent-1 ██████████ gates 6/6 confidence 0.94 ✓
agent-2 ██████░░░░ gates 4/6 confidence 0.41 ✗
agent-3 ████████░░ gates 5/6 confidence 0.67 ✗
agent-4 ██████████ gates 6/6 confidence 0.88
✓ agent-1 verified → merged to main (+147 -12)
Verification over hope.
Code proves itself through gates, not self-reported status.
Competition over iteration.
Why polish one attempt when four can race to the answer?
Evidence over opinions.
Every merge ships with diffs, logs, screenshots, and votes.
Fail closed, always.
If it can't prove it works, it doesn't ship. Period.
Spawn N agents on the same spec, each in an isolated git worktree. Different approaches, same acceptance criteria. The best code wins — ranked by evidence, not self-assessment.
Every candidate runs through independent gates: build, lint, test, visual diff. No gate pass, no merge — zero exceptions. High confidence auto-merges; anything below threshold queues for human review.
Multiple vision LLMs independently review before/after screenshots and vote. Catches CSS regressions, layout breaks, and visual bugs that unit tests can't see.
Model agnostic — any LLM via OpenRouter, local, or custom endpoints. Mix models in the same tournament. Drops into any CI pipeline with zero configuration.
Diffs, build logs, test output, before/after screenshots, council votes — every decision is auditable and archived automatically.
Gates passed
Council vote
Confidence
Above 0.9 → auto-merge
Artifacts
diff.patch
test-output.log
screenshot-before.png
screenshot-after.png
build.log
council-votes.json
confidence.md
spec.yaml
10×
faster feature delivery
0
unverified merges to main
93%
less time on code review
24/7
ships while you sleep
30-minute walkthrough. We'll run a live tournament on your codebase.