Loop Orchestrator
skill · operations · by Timan
The pattern behind Karpathy's autoresearch and John Ennis's Shoemaker's Elves — extracted and generalized for any AI agent work, not just ML training or coding. One-shot delegation fails for a predictable reason: you hand off a task, get back something that drifted from the brief, and have no structured way to catch it before the damage compounds. The Loop Orchestrator fixes that. Define the objective, assign an owner, run a bounded iteration, validate against a domain-specific metric, decide the next move. Repeat until done or until you kill it. The orchestrator keeps judgment. Specialists do substeps. The validation layer is where this earns its keep. Coding iterations check against tests, behavior, and pattern fit. Research iterations check against evidence quality and uncertainty reduction. Copy iterations check against voice, clarity, and goal alignment. Ops changes check against state change and no breakage. Strategy assumptions check against whether the assumption was actually answered and whether it changed a decision. Each domain gets a validator that matches what quality actually means there — not a generic rubric you'll ignore after round two. Batches run 30-45 minutes with drift detection built in. If work is moving away from the objective, you catch it at the batch boundary, not after you've shipped something wrong. Includes three templates: a survival guide for when iterations go sideways, an execution log for running decisions and outputs, and a review protocol for structured close-out at each loop. Built on Geoffrey Huntley's Ralph Loop, Karpathy's autoresearch pattern, and John Ennis's Elves framework. Tested on coding sprints, business validation experiments, and content production loops. Works solo or with multi-agent teams. Pairs directly with The Dev Team on souls.zip for full-stack agent workflows that actually finish.
Price: Free
Downloads: 9 · Version: 1.0.0
Rating: No reviews yet
API: GET /api/v2/items/loop-orchestrator
SkillOperations
Loop Orchestrator
Karpathy's autoresearch loop generalized beyond code. Bounded iterations for coding, research, copy, ops, strategy.
The pattern behind Karpathy's autoresearch and John Ennis's Shoemaker's Elves — extracted and generalized for any AI agent work, not just ML training or coding.
One-shot delegation fails for a predictable reason: you hand off a task, get back something that drifted from the brief, and have no structured way to catch it before the damage compounds. The Loop Orchestrator fixes that. Define the objective, assign an owner, run a bounded iteration, validate against a domain-specific metric, decide the next move. Repeat until done or until you kill it. The orchestrator keeps judgment. Specialists do substeps.
The validation layer is where this earns its keep. Coding iterations check against tests, behavior, and pattern fit. Research iterations check against evidence quality and uncertainty reduction. Copy iterations check against voice, clarity, and goal alignment. Ops changes check against state change and no breakage. Strategy assumptions check against whether the assumption was actually answered and whether it changed a decision. Each domain gets a validator that matches what quality actually means there — not a generic rubric you'll ignore after round two.
Batches run 30-45 minutes with drift detection built in. If work is moving away from the objective, you catch it at the batch boundary, not after you've shipped something wrong.
Includes three templates: a survival guide for when iterations go sideways, an execution log for running decisions and outputs, and a review protocol for structured close-out at each loop.
Built on Geoffrey Huntley's Ralph Loop, Karpathy's autoresearch pattern, and John Ennis's Elves framework. Tested on coding sprints, business validation experiments, and content production loops.
Works solo or with multi-agent teams. Pairs directly with The Dev Team on souls.zip for full-stack agent workflows that actually finish.
// security
Not reviewed
Not reviewedLast scanned Not available
Screening language is intentionally conservative. Review the bundle contents and capability needs before installing.
// source
Open source on GitHub
This item is sourced from a public GitHub repository. Install it here, or clone directly from GitHub.
// details
- Orchestrator-first — never delegate the whole loop, only bounded substeps
- Domain-matched validators for coding, research, writing, ops, and strategy
- 30-45 minute batch sizing with automatic drift detection
- Survival guide + execution log + review protocol templates included
- Works solo or with multi-agent teams (pairs great with The Dev Team)
- Built on the Ralph Loop pattern from Karpathy's autoresearch and Ennis's Elves
v1.0.1
Latest
1 weeks agoPending metadata update
v1.0.01 weeks ago
Linked from GitHub: https://github.com/heyastraio/loop-orchestrator
// creator
Timan
Startup founder. 15 years building and coaching startups. Sold my first company at 24. Coached 250+ teams from idea to validated business model. Now building heyAstra — a multi-agent team that productizes what I know. Astra handles coordination and strategy. Vega coaches health. More coming. Everything runs on OpenClaw, tested in production before it hits the shelf. I publish agents, skills, and workflows born from real problems. No theory pieces. If it's here, it ran the job first.
// creator
Timan
Startup founder. 15 years building and coaching startups. Sold my first company at 24. Coached 250+ teams from idea to validated business model. Now building heyAstra — a multi-agent team that productizes what I know. Astra handles coordination and strategy. Vega coaches health. More coming. Everything runs on OpenClaw, tested in production before it hits the shelf. I publish agents, skills, and workflows born from real problems. No theory pieces. If it's here, it ran the job first.
// related
Explore the neighborhood
Nearby items, alternatives, and more from the same creator.