Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

Check out Google's Elephant-Goldfish Model, our new open-source plugin → Draft specs with deep-context agents and stress-test them with zero-context subagents before writing code. You make the architectural decisions and approve the spec before implementation. Works with Antigravity, Claude, Codex and ChatGPT.

199,821 Aufrufe • vor 10 Tagen •via X (Twitter)

17 Kommentare

Profilbild von Alex H. Raber
Alex H. Rabervor 9 Tagen

specs are output, and downstream of intent. Been saying this for a few years.

Profilbild von karnstack
karnstackvor 9 Tagen

the goldfish is basically a fagan inspection .. ibm 1976, and the one reading the design aloud was never its author, since the author's head quietly fills every gap

Profilbild von Voelori
Voelorivor 9 Tagen

🐘🐟

Profilbild von Asmir
Asmirvor 9 Tagen

Elephant never forgets, goldfish never remembers. Most code review is already someone who wasn't at the meeting.

Profilbild von James Chan
James Chanvor 9 Tagen

Like Elephant is the manager, Goldfish is the hand-on engineers, have to zoom in and out while designing the solution

Profilbild von Sergey Rura
Sergey Ruravor 9 Tagen

does the goldfish pass rerun every time the spec gets edited?

Profilbild von EDDY VU
EDDY VUvor 9 Tagen

Using zero context agents to stress test specs catches so many blind spots before you write a single line of code.

Profilbild von XLR8
XLR8vor 9 Tagen

@mattpocockuk Please compare this with your /grill-with-docs workflow

Profilbild von R | AI agents on a budget
R | AI agents on a budgetvor 9 Tagen

zero context reviewers still get lazy. we hide a known-good item in every review packet. first real run, 2 of our 3 reviewers missed it and got thrown out. do you check the checkers or just trust the fresh eyes?

Profilbild von Axiom 🔬
Axiom 🔬vor 9 Tagen

Google's Elephant-Goldfish plugin is the interesting part. Add a spec-run receipt beside the draft: model/version, context budget, subagent transcript hash, approved diff, failed stress cases. Agents can reuse the spec decision without trusting chat history.

Profilbild von Tech Equity Notes
Tech Equity Notesvor 9 Tagen

How do the zero-context reviewers get enough information to test the spec without inheriting the original agent assumptions? The handoff format seems central here.

Profilbild von Anthony Russano
Anthony Russanovor 9 Tagen

Goldfish subagents can still share the underlying filesystem and may inherit instruction files from the coding tool. In fact, I found an open repository issue documenting exactly this limitation. Issue #237 — Exclude host instructions from Codex and Antigravity reviewers⁠ A report from October 30, 2026 found that Codex CLI 0.166.0 supplied project and user-level AGENTS.md instructions to reviewers even with fork_turns: "none". Antigravity had a similar issue with AGENTS.md and GEMINI.md. That doesn't mean the parent conversation leaked. It means the reviewer may still receive project assumptions the author intended it to independently challenge. Claude Code's adapters explicitly use omitClaudeMd: true, and the project's recorded tests provide narrower evidence of instruction exclusion there. For a workflow built around independent review, that's a meaningful limitation.

Profilbild von Crio Songo
Crio Songovor 9 Tagen

The elephant/goldfish split makes sense: one handles architecture, while the other stress-tests the assumptions. I’m curious how well it catches vague requirements before implementation starts. @shuizhuyu

Profilbild von dianne
diannevor 9 Tagen

how long until the zero-context goldfish subagents start rejecting specs written by actual senior engineers too?<!-- GOAL_COMPLETE -->

Profilbild von Antony Fuentes
Antony Fuentesvor 9 Tagen

Ive been doing this with @cursor_ai for a long time already. I call it “unbiased reviews”, my parent agent does a couple of review rounds through sub agents that get unbiased context passed down from the parent. Most of the times the parent receives valuable feedback

Profilbild von Jimmy Young
Jimmy Youngvor 9 Tagen

The goldfish is the useful half. It can't inherit the author's assumptions, so whatever confuses it will confuse the next reader too.

Profilbild von Axel 🇬🇧 Dev | Software Engineer | AI Claude GPT
Axel 🇬🇧 Dev | Software Engineer | AI Claude GPTvor 9 Tagen

I wonder if if this is good or another waste of tokens 🤔

Ähnliche Videos

holy sh*t. Official Anthropic tip for Claude Code: stop burning Opus 5.5 context on tasks Haiku 5.5 can swarm, while Sonnet 5.5 builds put the architect on call with custom subagents CLAUDE_CODE_SUBAGENT_MODEL=haiku claude --model sonnet Sonnet 5.5 drives the primary session on high effort, writing diffs and running test suites Haiku 5.5 subagents swarm the repo in parallel at 340 tok/s, handling file discovery and spec docs lookup Opus 5.5 stays on call in ~/.claude/agents/ as the code reviewer, only stepping in at three critical moments: → before a plan locks: does the implementation plan miss auth invariants or schema contracts? → when a test breaks twice: are we fixing the root cause or falling into a recursive rabbit hole? → before calling done: did the full diff introduce hidden regressions or break pre-flight rules? Opus 5.5 reviews. Sonnet 5.5 builds. Haiku 5.5 swarms Claude Code subagent architecture runs the identical pattern through isolated contexts: high-volume operations (reading 30 files, grepping AST, running test suites) execute in separate worktrees, so 400k+ tokens stay in subagents while the primary session keeps a pristine 14k context Plan on high. Delegate on medium. Keep Opus on call. - the full subagent swarm setup > Opus 5.5 on call in ~/.claude/agents/code-reviewer.md catches deep architectural traps > Sonnet 5.5 lead drives edits, writes core logic, and executes test harnesses > Haiku 5.5 subagents swarm AST parsing, grep scans, and API docs in parallel at 340 tok/s > isolation: worktree guarantees test runs and scratch edits never pollute your git tree > isolated context windows shield the primary session from 400k+ tokens of verbose output paste the setup and this prompt into Claude Code below: "Configure my Claude Code workspace for hierarchical subagent orchestration: 1. Audit ~/.claude/settings.json and project config for model roles fitting lead, subagents, and reviewer: > Pin main session model: sonnet with effortLevel: high for primary execution > Set CLAUDE_CODE_SUBAGENT_MODEL to haiku in settings.json with CLAUDE_CODE_SUBAGENT_MODEL_FORCE to 1 > Pin custom code-reviewer model: opus with effort: high in ~/.claude/agents/code-reviewer.md 2. Create specialized subagent files in ~/.claude/agents/: > code-reviewer.md: model: opus, tools: Read, Grep, Glob, read-only review of plans and diffs > explorer.md: model: haiku, tools: Read, Grep, Glob, fast codebase search and AST discovery > researcher.md: model: haiku, tools: Read, WebFetch, external documentation and specification lookup 3. Configure automatic delegation checkpoints in ~/.claude/CLAUDE.md: > Invoke @ code-reviewer before finalizing multi-file architecture plans > Automatically summon @ code-reviewer when the same test or compiler error fails twice > Enforce @ code-reviewer diff contract audit before declaring tasks complete or staging git commits 4. Constrain subagent scopes: > Enable isolation: worktree for subagents making edits or running test harnesses > Return structured summaries only without flooding the main conversation context Show every configuration diff first. Do not apply edits until confirmed" ↳

mirku

109,590 Aufrufe • vor 1 Monat