Two AI Coding Agents, One Token Budget

The fastest way to burn a Claude Opus budget is to make it do work that doesn't need a brain. Web searches, downloading datasets, grepping a 40k-line log, running a benchmark until it prints PASS — none of that needs Claude's reasoning, but all of it floods Claude's context window with tokens you pay for.

So I stopped. I now run two agents at once: Claude Code as the orchestrator that keeps the judgment, and Google's Antigravity CLI (agy) as the worker that eats the grunt work — on a separate quota. That's balanced development: the expensive model thinks, the cheap model fetches, and neither does the other's job.

The core idea: two fuel tanks

Claude Code and agy bill against different accounts. Claude runs on your Anthropic tokens. agy runs on your Google One / Antigravity quota. They don't share a meter.

That one fact changes how you budget a session. Every unit of bulky, low-judgment work you push to agy is tokens Claude never spends — and it comes out of a pool you're often not touching anyway. You're not shrinking one bill; you're moving work to the tank that's full.

agy is Google's official replacement for the Gemini CLI (which sunsets for individual tiers on 2026-06-18). It runs Gemini 3.x, Claude, and open-weight models, with web search, file edits, and shell built in. Headless mode is the part that matters here:

agy -p "find the latest stable pgvector version and output ONLY the version string" --model "Gemini 3.5 Flash (Low)"

That runs entirely inside agy. The search results, the page fetches, the noise — none of it touches Claude's context. Claude gets back one line: 0.8.0.

Orchestrator vs. worker: who does what

The split isn't "Gemini is worse, give it scraps." It's about what kind of value the task produces.

Give to agy (the worker)	Keep in Claude (the orchestrator)
web search → the answer	architecture and trade-off decisions
download files / datasets	security-sensitive code and review
grep huge logs/codebases → the hit	synthesizing the final result
run a validator/benchmark → PASS/FAIL	shipping anything without review
first-draft UI scaffold, boilerplate, tests	the final call on all of it

The rule I use: delegate when the cost is volume, or when the value is a reviewable draft. Keep it when the value is a final decision.

That second clause is the interesting one. agy isn't only for legwork — Gemini 3.5 Flash is genuinely strong at first-draft frontend and boilerplate. So I let it scaffold a component or generate a test file on its quota, and Claude reviews a diff instead of starting from a blank page. A draft from the worker is an input to Claude's judgment, never a substitute for it.

A real session, balanced

Here's a "benchmark my system against competitors" task split across both agents:

1. Claude plans the benchmark: what to measure, which competitors, what counts as a win.
2. agy does the legwork — searches each competitor's docs, downloads sample outputs, greps my logs for latency numbers. All that I/O stays in agy's context.
3. agy writes raw results to disk in write mode, then replies with just the file path.
4. Claude reads only the slice it needs from that file and writes the conclusion.

The pattern is file-out, read-the-slice: bulky output lands on disk, and Claude reads only what it has to reason about. Claude never ingests a raw search dump or a megabyte of downloaded data — yet it makes every decision. Same instinct behind a zero-cost AI stack: route the cheap, high-volume work to where it's free.

Picking the worker's model

agy models lists what your account can use. I route by subtask and pay only for the effort I need:

Subtask	Model
search / download / scan / validate	Gemini 3.5 Flash (Low) — cheapest, fastest
frontend/UI drafts, code-gen	Gemini 3.5 Flash (High)
deep reasoning, long context	Gemini 3.1 Pro (High)
correctness-sensitive code, test gen	Claude Sonnet 4.6 (Thinking)
hardest reasoning	Claude Opus 4.6 (Thinking)

Default is Gemini 3.5 Flash (Low). Cheapest-capable wins; escalate only when the worker itself has to reason.

Why this is "balanced," not just cheap

Token efficiency is the headline, but the deeper win is separation of concerns between models. Claude's context stays clean — it holds the plan and the decisions, not the byproducts of fetching them. On long sessions that keeps its reasoning sharp, because the window isn't half-full of log spam. Cheaper and clearer at once.

The discipline that makes it safe: never ship a delegated result blindly. Verify the worker's answer; review the worker's draft. The doing is delegated; the judgment never is. It's the same orchestration mindset I lean on as a solo founder running on Claude Code, and it pairs with the rest of my tools for AI coding agents.

I packaged this workflow as an open-source agent skill so Claude Code (and Cursor, Windsurf, Codex, Cline) can apply the orchestrator/worker split automatically. Install agy first, then the skill.

Frequently Asked Questions

Does delegating to agy cost Claude tokens? No — agy runs on your Google One / Antigravity quota, a separate meter from Claude's. Only the small distilled answer it returns enters Claude's context.

Won't a cheaper model produce worse work? For legwork (search, download, scan) quality is binary — you got the version string or you didn't. For drafts, Claude reviews before anything ships, so the floor is set by Claude, not the worker.

What is agy? The Antigravity CLI — Google's official replacement for the Gemini CLI. It runs Gemini 3.x, Claude, and open-weight models headlessly, with web search, file edits, and shell.

When should I NOT delegate? Architecture decisions, security review, anything tiny (delegation overhead isn't worth a few lines), and shipping code without review. Keep judgment in Claude.

---

Written by Shihab Shahriar Antor — AI Engineer & Founder of Shahriar Labs. The orchestrator/worker skill is open source: antigravity-delegate on GitHub.