BIOME: Evolving Google's Elephant-Goldfish Architecture for Production Multi-Agent Engineering

makari — Mon, 09 Nov 2026 11:21:42 +0000

Building reliable software with LLM agents is notoriously difficult. Single-prompt coding agents degrade quickly, suffer from context rot, and hallucinate across large codebases.

Recently, Dave Rensin and the Google Cloud team introduced a brilliant breakthrough: The Elephant-Goldfish Model (EGM). It introduced the separation of concerns between:

The Elephant: Long-term conversational memory, architectural context, and specification drafting (strictly forbidden from modifying source code).
The Goldfish: Stateless, zero-context reviewers that independently stress-test design specs and inspect diffs with fresh eyes.

We loved this paradigm, but when putting pure EGM into production on multi-stack repositories (Database + Backend API + Web + Mobile), we ran into real-world operational bottlenecks.

Today, we're releasing BIOME (Apache 2.0) — a production-grade evolution of EGM that tackles these hard engineering challenges.

The Engineering Gaps & How BIOME Solves Them

1. The Elephant Context Bloat

In massive repositories, having the Elephant read source code directly quickly exhausts context windows and triggers hallucinations. Keeping it 100% blind, however, leads to inaccurate interface specs.

Solution: The Falcon & The Cartographer

The Falcon: An ephemeral reconnaissance scout that inspects signatures, entrypoints, and schemas, returning only lightweight summaries to the Elephant.
The Cartographer: Audits codebase topology once and caches it in .biome/codebase_topology.json, eliminating redundant repository-wide scans.

2. The Monolithic Builder Bottleneck

Original models relied on a single generalist builder agent. When modifying complex full-stack apps, a single builder modifying DB schemas, backend controllers, and mobile screens in one pass introduces subtle cross-layer drift.

Solution: The Beaver & Domain Pods
We introduced The Beaver — surgical, ephemeral builders partitioned into strict domain pods:

db: Schema migrations and models
backend: Core business logic, APIs, and service layers
web: Frontend interfaces
mobile: Client applications
e2e: Integration verification

Each pod executes in isolation and must pass local domain tests before the next pod begins.

3. Context Compaction Amnesia

When working on large tasks, host harnesses (like Claude Code, Antigravity, or Copilot) automatically summarize or compact conversational history. In standard agents, this wipes out the task state, causing the coordinator to lose track and start writing unreviewed code.

Solution: The Elephant State Anchor
BIOME maintains a deterministic state file at .biome/state.json. Upon any compaction event, the post-compaction recovery protocol immediately forces the Elephant to reload its state anchor and resume from active_phase without missing a beat.

4. Eradicating "AI Slop"

LLM-generated code often suffers from robotic anti-patterns: bloated class names (*ManagementProcessorHelper), chatty redundant comments, and synthetic UI text.

Solution: The Chameleon & Woodpecker Vector 5

The Chameleon: An anti-slop stylist enforcing idiomatic human code style, concise naming, and clean microcopy.
The Woodpecker: Relentlessly pecks at git diffs across 5 strict vectors: Security, Specification Fidelity, Edge-case Regressions, Micro-Architecture, and Anti-Slop.

5. Zero-Zombie Lifecycle Hygiene

Headless subagents can linger in background tasks, leaking resources and polluting git state. We implemented The Reaper, which strictly enforces a "Zero Zombie" invariant by terminating all subagent sessions immediately upon task completion.

The Animals of BIOME at a Glance

Animal	Context Mode	Responsibility
🐘 The Elephant	Project context & specs	Explores problems, drafts specs, coordinates delivery. Never writes code.
🦅 The Falcon	Rapid reconnaissance	Probes entrypoints and signatures without polluting Elephant context.
🐠 The Goldfish	Stateless review	Stress-tests assumptions and exposes edge-case blind spots in specs.
🗺️ The Cartographer	Topology engine	Caches module relationships into `.biome/codebase_topology.json`.
🦫 The Beaver	Ephemeral builder	Executes surgical mutations across partitioned domain pods.
🦎 The Chameleon	Anti-slop stylist	Enforces natural human naming and cleans robotic AI boilerplate.
🪵 The Woodpecker	Zero-context diff inspector	Audits code mutations across 5 rigorous vectors.
⚡ The Reaper	Lifecycle protocol	Purges worker subagent sessions immediately upon task completion.

Status & Open Source

Test Suite: 340+ unit and integration tests passing.
Runtimes: Pre-packaged distributions for Claude Code (distributions/claude) and OpenAI / OpenCode environments.
License: Apache 2.0.

Huge thanks to Dave Rensin and Google for pioneering the conceptual model that made this possible.

Check out the repository, try it out, and let us know what you think:
👉 GitHub: https://github.com/asteria-core/biome

DEV Community: makari