The Four Loop Patterns in Claude Code
Turn-based, goal-based, time-based, proactive. What each loop pattern is, when to reach for it, and the three ingredients that make a loop actually finish: done, budget, checkpoint.
Claude Code orchestrates; Antigravity's agy CLI does the token-heavy legwork on a separate Google quota. The orchestrator-worker split, explained.
The cheapest way to run an AI coding agent is to stop making the expensive one do cheap work. At Shahriar Labs, day-to-day development runs two agents at once: Claude Code as the orchestrator that holds the judgment, and Google's Antigravity CLI (agy) as the worker that absorbs the token-heavy legwork — on a separate quota.
Orchestrator–worker split (Noun): a development pattern where a high-capability model (Claude Code) plans and decides, while a separate, cheaper agent (agy) executes high-volume, low-judgment subtasks — search, downloads, scans, validators, first-draft code — and returns only the distilled result.
Claude Code bills against Anthropic tokens; agy bills against a Google One / Antigravity quota. They do not share a meter. Every unit of bulky work pushed to agy is tokens Claude never spends — drawn from a pool that is usually idle. This is "balanced development": the expensive model thinks, the cheap model fetches, and neither does the other's job.
The boundary is the kind of value a task produces, not raw capability. Volume work and reviewable first drafts go to the worker; final decisions stay with the orchestrator.
agy: web search, downloading datasets, grepping huge logs, running benchmarks to PASS/FAIL, and first-draft UI scaffolds, boilerplate, and tests.The rule: delegate when the cost is volume or the value is a reviewable draft; keep it when the value is a final decision. A draft from agy is an input to Claude's judgment, never a substitute for it.
agy models lists what an account can use; Shahriar Labs routes by subtask and pays only for the effort needed. Gemini 3.5 Flash (Low) handles search, downloads, scans, and validation — cheapest and fastest. Gemini 3.5 Flash (High) handles frontend and code-gen drafts. Reasoning-heavy subtasks escalate to Gemini 3.1 Pro or a Claude (Thinking) variant. Cheapest-capable wins.
For bulky output, the worker writes results to disk and replies with just the path. Claude then reads only the slice it must reason about — never a raw search dump or a megabyte of downloaded data — yet it still makes every decision. The same instinct powers products like LetX: push high-volume work to where it is cheap, keep the judgment where it matters.
Q: Does delegating to agy cost Claude tokens?
A: No. agy runs on a Google One / Antigravity quota, separate from Claude's tokens. Only the distilled answer enters Claude's context.
Q: Won't a cheaper model produce worse work?
A: Legwork quality is binary, and every draft is reviewed by Claude before shipping — so the worker never sets the quality floor.
Q: Where can I read the full engineering write-up?
A: The founder's deep dive is on the hub at shihub.online, and the workflow is open-sourced as antigravity-delegate on GitHub.
Pairing Claude Code with Antigravity turns one strained token budget into two balanced fuel tanks. The orchestrator reasons; the worker fetches and drafts on a separate quota; nothing ships without review. Cheaper, cleaner, and faster at once.
Written by Shihab Shahriar Antor — AI Engineer & Founder of Shahriar Labs. Building LetX, QuantumSketch, and open-source AI agent skills.
Turn-based, goal-based, time-based, proactive. What each loop pattern is, when to reach for it, and the three ingredients that make a loop actually finish: done, budget, checkpoint.
How we ported 5GB of TeX Live to the browser, and why WebAssembly is the future of complex web apps.