Product.ai / Join / Projects / Sub-Agent Orchestration with Worktree Isolation — refactor one Cortex or AIOS workflow into the engineer-plus-agent-fleet pattern
Project Open to Alpha Team

Sub-Agent Orchestration with Worktree Isolation — refactor one Cortex or AIOS workflow into the engineer-plus-agent-fleet pattern

Pick one Cortex or AIOS workflow currently running sequentially (or with naive sub-agent dispatch) and refactor it into the engineer-plus-agent-fleet pattern. Multiple sub-agents in parallel, each in its own git worktree for spatial isolation, with structurally separated information bases (oracle separation between generator and verifier), explicit external truth anchors between phases, and a deterministic synthesis gate at the end. Real workflow, real commits, measured speedup, measured quality delta.
Project Overview
Discipline
AI Systems — AI Engineer · ai-systems-engineer · ai-systems-associate
Duration
2 weeks
Compensation
Your stated freelance rate
Surface
Engineering · Cortex · Product.ai
Kernels
engineering · cortex · productai
Outcomes
team-velocity · team-visible · dev-integrate
Tier
Consequential
Alpha Team
Open to alpha members who want to take this on
Tooling
Claude Code or Co-work

Why we want this done

Senior engineers at frontier firms run 5-15 simultaneous agent sessions with specialized roles. Steinberger ships 118 commits/day with 5-10 concurrent agents. The bottleneck moved from typing to scoping, context-providing, reviewing, and drift-detection. The pattern is engineer-driven parallelism with deterministic isolation (git worktrees), NOT multi-agent system orchestration with stateful coordination protocols (which fails 40% within 6 months — Multi-Agent Physics Meta-Axiom). Cortex and AIOS today have several skills running sequential sub-agent fan-outs without spatial isolation, oracle separation, or external truth anchors between phases. This is exactly the closed-loop compound-deception failure pattern documented in the Anthropic dual-postmortem corpus. The candidate refactors one workflow into the structurally correct pattern, demonstrates the speedup empirically, and produces the orchestration primitive other workflows adopt.

Scope

  1. Audit candidate workflows — Cortex /cortex-update, AIOS /share, ARC autopilot phases, signal-step-executor, or another sub-agent fan-out path
  2. Pick one (the candidate proposes; we pressure-test)
  3. Map the current architecture — what runs sequentially, what could run in parallel, what passes through closed-loop verification
  4. Refactor: spawn sub-agents in worktrees (Agent tool isolation: "worktree" parameter), oracle-separate generator from verifier, add named external truth anchors between phases
  5. Build a deterministic synthesis gate — the merge step is invariant-checked, not model-judged
  6. Measure: end-to-end time, token spend, quality delta against a frozen golden set
  7. Document the orchestration primitive — when to fan out, when not to, how to scope worktrees, how to wire the truth anchors

What success looks like

  • The workflow runs measurably faster end-to-end (target: ≥2x on parallelizable phases)
  • Quality is preserved or improved against the frozen golden set
  • Each sub-agent runs in its own worktree (or equivalent spatial isolation primitive); no shared mutable state across the fan-out
  • The generator and verifier sessions are oracle-separated — the test-generating session does not see the implementation, and vice versa
  • An external truth anchor is nameable in one sentence per phase — a check the agents cannot influence
  • The orchestration primitive lives in a reusable form (a skill helper, an Agent-tool wrapper, or documented harness pattern)
  • The Multi-Agent Physics 24-hour archaeology ritual fires on any near-miss the workflow produces during the trial

References

references.md
AI Engineering Phase 3 briefing axiom E1 (Engineer-Plus-Agent-Fleet), F2 (Substrate-vs-Harness Boundary), F4 (Phase-Sequenced Adoption)
Cortex CLAUDE.md §5.2 rules 7-10 (Multi-Agent Engineering Physics)
axioms/Engineering/Multi-Agent Physics (2026-04-18).md (full physics)
aios/.claude/skills/ for existing fan-out patterns
Steinberger 5-10 concurrent agents pattern (citation in axiom E1)
ARC Autopilot pipeline (if accessible — note IP-separation if not) for fan-out scaffolding
Cursor 1000-agent browser project + Anthropic 16-agent C-compiler (cited in CLAUDE.md §5.2 rule 8)

Constraints

  • Claude Code as primary substrate (Agent tool with isolation: "worktree" is the load-bearing primitive)
  • Probabilistic discipline alone does not satisfy the brief — every shared-surface write must be paired with a deterministic invariant check (CLAUDE.md §5.2 rule 7)
  • Atomic commits — workflows that touch HTML+JS pairs must respect the rule 3 hard subrule
  • The 24-hour archaeology ritual (rule 11) applies to any failure during the trial — race-class taxonomy entry + DEPENDENCY-PAIRS.md update + shipped enforcement gate
  • IP separation: ARC Autopilot and other aios-methods paths are out of scope; pick a workflow that lives in the application layer
  • Quality-first cost-aware: parallelization that doubles spend without doubling output is a fail
Apply
01

Read the Codex (10 min)

The operating principles we work by. If they resonate, the rest of this will land. Open the Codex →

02

12-minute video screen

Hireflix, async. Questions are calibrated to this project specifically.

03

Chemistry call (30-60 min)

Direct call with the CEO. Strategic alignment and mutual fit. No problem-solving exercise.

04

Project begins within 2-3 weeks

1099 contractor agreement, NDA, paid at your stated rate. Day 1 in Santa Monica.

Alpha Team members can take this project without the screen-and-call sequence. Reach out via the Alpha Team channel.

Apply for this trial Browse other projects