Amiga IA
Comprehensive ecosystem of 14 autonomous subagents, stateless guardrail hooks, and 25 portable declarative skills. 100% cross-compatible with Antigravity (Gemini), Claude Code, and OpenAI Codex.
The Problem
Traditional step-by-step imperative instructions break down as codebases scale, alongside heavy recurring token taxes caused by bloated system prompts and repetitive multi-engine context.
The Solution
A unified agentic ecosystem featuring 14 autonomous orchestrating subagents and 25 declarative skills, powered by a tri-ecosystem compiler (Antigravity, Claude Code, and OpenAI Codex), delimited block distribution, Double-Blind Cleanroom TDD, and deterministic empirical benchmarks.
Key Learnings
- Empirical Benchmark Validation (ADR-004, ADR-012): Offline deterministic benchmark suite demonstrating a 44.75% (-2,973 tokens) reduction in context overhead per turn, 0% confirmation bias in Cleanroom TDD, and 100% precision and recall (F1 = 1.00) in repository defect audits.
- Idempotent AI Rules Distribution: Designed a delimited block protocol to inject and sync agent directives across codebases without overwriting local configurations or leaving artifacts.
- Release Automation Verification Gates: Enforced live artifact verification protocols ensuring compiled binaries are physically generated and uploaded before publishing official releases.
- Agentic vs. Passive Prompting: Transitioning to autonomous subagent profiles that reason about project goals and deploy worker threads in parallel proved dramatically more robust and scalable than monolithic prompt engineering.
- Tri-Ecosystem Semantic Compiler: Automated compilation pipeline transforming canonical Markdown sources into native structures for Anthropic Claude Code (~/.claude/), Google Antigravity (~/.gemini/config/), and OpenAI Codex (~/.agents/ and ~/.codex/).
Technical Architecture & Orchestration
Amiga IA elevates AI-assisted software engineering from linear chat prompts to an autonomous subagent mesh of 14 specialized agents (ami-tech-lead, ami-repo-auditor, ami-push-assistant, etc.) supported by 25 on-demand declarative skills. Each agent operates with bounded tools, declarative skills, and explicit verification gates.
- Universal Adapter & Multi-Engine Portability: Stateless guardrail injection and portable skill compilation with native cross-compatibility across Antigravity (Gemini), Claude Code, and OpenAI Codex.
- Delimited Block Protocol: Idempotent synchronization engine that updates global rules across codebases without overwriting local configurations or leaving residual clutter.
- Double-Blind Cleanroom Isolation: Decoupled TDD workflow (
ami-cleanroom-arbiter) where implementers and test authors communicate exclusively over frozen interface contracts, eliminating confirmation bias.
Optimization, Benchmarks & Verification Gates
- Empirical Benchmark Suite: Offline deterministic evaluations demonstrating a 44.75% (-2,973 tokens/turn) reduction in context overhead, 0% escaped defects under Cleanroom TDD, and perfect precision and recall ($F_1 = 1.00$) in repository defect detection.
- Anti-Hallucination Release Protocol: Automated release pipelines with live artifact verification gates that compile, test, and verify binary checksums before enabling publishing to GitHub.