Orchestrate Parallel Code Generation Agents

This skill coordinates quota-aware, parallel subagent missions for Antigravity 2.0, turning one big task into isolated agent missions without burning weekly quota. It runs six phases. Decompose: before spawning anything, it announces "Running subagent-orchestrator skill" and produces a Mission Brief (goal, total agents, quota strategy, expected token cost, and per-agent ID/role/scope/model/input/output/dependencies) that the user must approve before proceeding. Quota routing: a decision tree assigns models - tasks over 20 files or 500 new lines get Gemini Flash for every agent with Sonnet reserved for final review; creative UI or complex-logic/API-design work gets Sonnet for the builder agent and Flash elsewhere; everything else gets Flash for all agents - under hard cost rules that Claude Opus is never used in subagents, Claude Sonnet gets at most one subagent per mission, Gemini Flash is the default, and a browser subagent runs on its own pool at most once per mission. Context isolation: each subagent gets a scoped context packet naming exactly which files to read, which to write, which to explicitly exclude, and only the relevant section of the project's knowledge file, with irrelevant directories (node_modules, package-lock.json, .next/, dist/) added to .antigravityignore.

Use it when a task spans 3+ files or components, multiple agents should work simultaneously, quota issues have hit mid-task before, the task involves both planning and building, or a browser agent, code agent, and terminal agent need to run together. It explicitly does not apply to editing a single file, fixing one bug, writing a quick script under 50 lines, or a question/plan-only request.

Parallel execution runs agents in dependency rounds (no-dependency agents in parallel, dependent agents after their inputs land, a final integration round), with a mandatory 3-point spot check between rounds - did each agent stay in scope, are there import/export conflicts, did any agent leave a placeholder like "TODO" - re-running any agent that fails a check rather than continuing. Error recovery never re-runs the full mission: it identifies the exact failure point and spawns a single repair agent scoped to only the broken files, with the error message as context and Gemini Flash as the cheapest repair model, validating before continuing. A final integration check verifies imports resolve, no duplicate function/variable names exist across files, no hardcoded values that should be env variables, no leftover console.log, consistent TypeScript types, and a mentally-verified build, spawning one more scoped repair agent if any item fails. Quota is tracked by event (file index and agent spawn are low impact; tool calls and terminal commands are medium; a browser subagent is high; thinking mode is very high) - crossing 60% of sprint quota mid-mission triggers a checkpoint report, a switch to Flash for remaining agents, and disabling any not-yet-started browser subagent.

Ships worked examples in an examples/ folder: a Next.js feature built with 3 parallel agents, a backend-API-agent-plus-frontend-UI-agent pair running in parallel, and a minimal-quota repair mission for a broken build. Communication rules require announcing which agent is running at all times, a compact progress bar between rounds showing both mission progress and quota status, and never going silent for more than one agent turn.

Antigravity 2.0 users running multi-file, multi-agent tasks who need explicit quota budgeting, scoped context per agent, and a structured recovery path - not a runtime scheduler or automatic quota enforcer, but a coordination discipline the parent agent still has to apply.