Treating every prompt as a flagship problem wastes weekly quota on boilerplate. Routing architecture or review to a lightweight tier trades catch-rate for cheap tokens and forces rework.
As of November 2026, we route foundry work across two cloud quota pools, three host subagent tiers, and a local LAN offload.
1. Quota Pools & Host Dispatch
Antigravity 2.0 splits cloud capacity into two rate-limit pools alongside local inference and host-managed subagents:
| Surface | Active Models / Tiers | Primary Role |
|---|---|---|
| Pool A (Google) | Gemini 4 Hydrogen (Low–XHigh), Carbon (Low–High), Argon (Low–High), Gemini 3.8 Flash (Low–High), Orchestrator (Carbon/Argon) | Daily coding (Hydrogen), frontier audits (Carbon), utility (Flash). |
| Pool B (Partner) | Claude Opus 5.5 (Low–High), Claude Sonnet 5.5 (Low–High), GPT-OSS 120B (Medium) | Cross-pool audits (Sonnet/Opus), forensics, and fallback. |
| Local LAN | Self-hosted 3B / 7B coder |
Zero-cloud-quota diff linting and commit synthesis. |
| Host Subagents | flash_lite, flash, pro (disallowed: inherit) |
Scoped child workers dispatched by a session. |
Omitted Model and inherit are banned across all tiers (Low–Very High) so child turns never clone parent quota.
Low/Medium tasks limit workers to flash_lite/flash and retire flash_lite from critics (flash minimum for Low/Med sub-issue reviews). High/Very High unlock pro for research and coaching; High+, Epic, and trunk reviews move to clean-room sessions.
2. Per-Model Opportunity Costs
| Model / Route | Best Fit | Sober Opportunity Cost |
|---|---|---|
| Gemini 4 Hydrogen (Low–XHigh) | Daily coding across Low–High | Shares Pool A with Carbon; over-efforting trivial edits drains audit quota. |
| Gemini 4 Carbon (Low–High) | Frontier audits; V.High / Epic | Scarce Pool A capacity; routine coding starves clean-room reviews. |
| Gemini 4 Argon (Low–High) | Legacy / non-audit utility | Trails high-effort Hydrogen; barred from clean-room audit stamps. |
| Orchestrator (Carbon / Argon) | Dynamic multi-step routing | Obscures model attribution; excluded from commit and audit stamps. |
| Gemini 3.8 Flash (Low–High) | Browser, lookups, Low/Med critics | Fast Pool A burn; misses cross-repo invariants without tight specs. |
| Claude Sonnet 5.5 (Low–High) | Cross-pool audits and fallback | Spares Pool A while burning Pool B; needs cold-start handoff context. |
| Claude Opus 5.5 (Low–High) | Deep forensic root-cause | Highest Pool B cost and latency; kept out-of-band, never routine PRs. |
| GPT-OSS 120B (Medium) | Pool B open-weights fallback | Fixed Medium effort; lacks headroom for High+ adversarial audits. |
Subagent flash_lite / pro |
flash_lite: status; pro: High+ research |
flash_lite drifts on ambiguity; pro adds high latency and burn. |
Local LAN (3B / 7B) |
Diff lint (3B/7B), commit text |
Hardware latency, ~4K context cap, and offline fallback if down. |
Clean-room Author != Reviewer pairing bars *-low audits and enforces an Authoring Boundary.
Audits run intra-Pool-A (Hydrogen -> Carbon, or Carbon -> Carbon when reviewer effort >= author effort, sharing Pool A) or cross-pool (Pool A -> Sonnet/Opus, spending Pool B to spare Pool A).
3. Pros & Cons
Pros
- Quota resilience: Two cloud pools plus local LAN offload prevent one rate limit from stalling work.
- Controlled subagent burn: Banning
inheriton all tiers and retiringflash_litecritics protects quota and catch-rate. - Uncorrelated reviews: Intra-Pool-A (
Hydrogen -> Carbon) and cross-pool (Pool A -> Sonnet/Opus) pairing catch blind spots.
Cons
- Handoff & pool contention: Cross-pool switches need cold-start summaries;
Hydrogen -> Carbonshares Pool A. - Misrouted complexity: Ambiguous tasks on
flash/flash_liteor diffs past ~4K local context cause rework or clipping. - Snapshot decay: Model rosters, effort bands, and pool limits shift faster than static routing tables.