AI agents are non-deterministic. Your orchestration shouldn't be.

Deterministic structure.
Non-deterministic intelligence.

Jupiter splits orchestration into two layers: Rust decides the structure, Claude decides the code. Same task, same skeleton, every time.

Private Alpha · Rust engine · 9/9 workers · 4/4 phases

jupiter — ~/project
$ jupiter run "Build e-commerce: auth, catalog, cart, checkout"
Analyzer — AI extracts features (JSON only)
→ ["auth-system", "catalog", "cart", "checkout"]
Archetype — Rust builds dependency graph (zero AI)
Phase 1: setup (1 worker)
Phase 2: 4 features (parallel, isolated worktrees)
Phase 3: integrator (1 worker)
Phase 4: tester + qa (parallel)
Executing Phase 2 — git worktree isolation
[W1] auth-system
done ✓
[W2] catalog
done ✓
[W3] cart
80%
[W4] checkout
60%
Build gate: cargo check pass
9/9 workers completed — 146 unit tests + 23 E2E passing

The architecture problem

AI is brilliant at code. Terrible at project management.

So Jupiter doesn't ask AI to be a project manager.

Today’s AI tools
With Jupiter
Run the same task twice, get two different outputs
Same task, same skeleton, every time
AI forgets workers, skips QA, duplicates features
Structure is hardcoded in Rust — QA always exists
8 parallel AI sessions just to generate briefings
1 AI session. 7/9 briefings are Rust templates
70KB bloated briefings that inline entire contracts
~500 token briefings, surgical and typed
No guardrails — broken phase 2 launches phase 3
Build gates between every phase. Fails → blocks
Parallel agents overwrite each other's files
Git worktree isolation — physically can't conflict

Analyze. Structure. Execute.

AI fills in the content. Rust decides the structure.

01

Analyze

AI (constrained)

A short Claude session (~500 tokens) extracts independent features from your task. Output: a JSON array. Nothing else.

02

Structure

Rust (deterministic)

Jupiter generates the entire dependency graph in Rust. Phases, workers, file ownership, briefings. No AI involved.

03

Execute

Rust orchestration

Workers run in isolated git worktrees. Build gates between phases. Failures re-spawn with context. Merge is sequential.

“Jupiter doesn’t replace AI creativity.
It contains it.”

The AI didn’t get smarter between v4 and v5. The orchestration did.

9/9
workers completed
4/4
phases completed
146
unit tests passing
56 min
total execution

What’s inside

Every property is enforced by Rust, not by prompting.

Archetype Engine

Rust generates the entire dependency graph deterministically. Phases, worker assignments, file ownership. Zero AI involved. The AI cannot skip a phase, forget QA, or create an invalid graph.

Git Worktree Isolation

Each parallel worker gets its own git worktree. Workers physically cannot modify each other's files. Merge is sequential, controlled by Rust.

Build Validation Gates

After each phase, Rust runs cargo check / tsc. If it fails, the next phase doesn't start. Circuit breaker resets worktree, re-spawns with stderr as context.

Rust-Templated Briefings

7 out of 9 workers get instructions without touching the AI. ~500 token briefings instead of 70KB. From 8 parallel Claude sessions to 1.

Private Alpha

The AI didn’t get smarter.
The orchestration did.

We went from 0/10 workers completing to 9/9 by replacing non-deterministic AI orchestration with a deterministic Rust engine. Join us and help shape what comes next.

Join the Waitlist on Discord

Feedback-driven · Early access · #JupiterAI