Gemini 3.5 Flash & Google Antigravity — Limitations and Execution Constraints

Status: Execution-target reference (not a feature design). Why this exists: the 2026-06-18 implementation plans are written to be executed autonomously by Gemini 3.5 Flash inside Antigravity. This doc records why those plans are shaped as they are, so future sessions targeting the same stack reuse the constraints instead of rediscovering them. Confidence grading: ★★★ official/primary · ★★ reputable practitioner report · ★ inferred.

1. What the stack is ★★★

Gemini 3.5 Flash — the fast model Google co-optimized for the Antigravity agent harness (not a general model shoehorned in). (Antigravity overview)
Google Antigravity (2.0) — an agent-first IDE (VS Code-based) whose orchestrator decomposes a goal and spawns dynamic subagents with isolated context windows, in parallel (demoed at 93 parallel subagents). Subagent capabilities are also exposed via a Managed Agents API. (Antigravity agent docs, DataCamp)

2. Model capability profile ★★★

Context: 1M-token input, 64k-token output. (Gemini 3 dev guide)
Strengths: GA, strong at agentic execution, coding, long-horizon tasks. (model card)
Weaknesses (vs 3.1 Pro): trails on Humanity’s Last Exam, ARC-AGI-2, and the 128K MRCR v2 long-context retrieval test — “not a clean replacement for Pro.” → weaker deep reasoning and weaker long-context recall. (Appwrite deep-dive)
Feature gaps: Computer Use not supported in 3.5 Flash (use 3 Flash Preview). (what’s new)

3. Documented reliability failure modes ★★ (practitioner reports)

Low real-world completion: ~48% task completion in-IDE vs ~80% benchmark — a large reliability gap. (3-month failure analysis)
“Agent execution terminated due to error” mid-run, no checkpoint, leaving files partially edited / non-compiling. (Agent-Terminated crisis)
Quota = binary hard cutoff, no warning — execution stops immediately, can leave a broken state.
Hallucination cascades: invents APIs / phantom libraries with confidence; serialized session state lets an early wrong assumption become “ground truth” in later sessions. (hallucination report)
Poor self-correction: repeats the same failed action instead of recovering.

4. Customization surface ★★

Reads GEMINI.md (Antigravity-specific, highest priority) and AGENTS.md (shared across Antigravity/Cursor/Claude Code, since v1.20.3 / 2026-03-05).
Skills are mounted from .agents/skills/ into the sandbox; agent personas via agents.md. (Antigravity rules guide, agents.md codelab)
Implication: it cannot see Claude’s external ~/.claude/ skill cache — any referenced skill must be in-repo (Vox now ships them under crates/vox-skills/skills/superpowers/).

5. Plan-engineering constraints (the rules implementation plans MUST follow)

Failure mode	Required plan constraint
No-checkpoint mid-task termination → broken tree	Every task is atomic and ends GREEN + committed. A kill between tasks leaves a compiling, tested tree. Never split a compile-breaking change across two commits.
Hallucinated APIs / phantom symbols	Verify-before-use: an explicit `rg`/read step precedes any code step that references a symbol/path; inline exact signatures. Forbid “assume the API.”
Weak long-context recall (MRCR)	Self-contained tasks: repeat needed context/code in each task; never rely on remembering earlier tasks.
Poor self-correction / repeats failures	Two-strike circuit breaker: verification fails twice → STOP + handoff note; do not loop.
Weak deep reasoning	One decision per step, explicit instructions; no open-ended “design X” steps in execution plans.
Quota hard cutoff	Small tasks + frequent commits so a cutoff wastes ≤ one task.
Isolated-context parallel subagents	Tag tasks PARALLEL-SAFE / SEQUENTIAL by file-write disjointness; never two subagents on one file.

6. Handoff prerequisites

AGENTS.md present; generate GEMINI.md (e.g. from CLAUDE.md) for Antigravity-specific overrides.
In-repo skills available under crates/vox-skills/skills/superpowers/ and referenced by the plans (see ../contributors/antigravity-handoff-and-skill-gaps-2026-06-18.md).
Run plan Pre-flight rg commands first (anti-hallucination baseline).

Sources

Antigravity overview/agent docs/DataCamp · Gemini 3 dev guide / 3.5 Flash model card / Appwrite deep-dive / what’s-new · failure-analysis, Agent-Terminated, hallucination reports · Antigravity rules guide / agents.md codelab — full URLs inline above.