Our team built a playable 3D Wizard Chess game from scratch with Gemini 4 Argon using Google Antigravity in the Gemini app for Mac.
Coordinate a multi-agent Blender workflow to model, design, and deliver fluid, real-time animations so you go straight from idea to play.
THIS IS AN ABSOLUTE F*CKING GOLD
Anthropic just revealed one of the sharpest setups for Claude Code
The idea: run autonomous agent teams across Opus 5.5, Sonnet 5.5, and Haiku 5.5, while keeping Fable 5.1 on call as adversary
Zero coordinator bottlenecks and way fewer tokens: teammates talk directly over peer-to-peer mailboxes and claim tasks on their own
You just launch Claude Code like this:
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude --model opus
here's how it splits:
→ Opus 5.5 acts as team lead: breaks epics into a shared task list, arbitrates conflicts, and signs off on pull requests
→ Sonnet 5.5 powers builder teammates: hammers out frontend and backend modules in parallel, messaging peer-to-peer without coordinator drag
→ Haiku 5.5 swarms lightweight scouts: maps AST call sites, audits dependencies, and verifies API specs at 340 tok/s
→ Fable 5.1 monitors the peer bus on call: audits cross-session messages and catches subtle edge cases before code merges
→ you stop burning Opus context on routine implementation loops and keep the lead session lean
Fable 5.1 only steps in at three moments:
→ before a contract locks: does the frontend payload match the backend schema?
→ when a test breaks twice: are we fixing the root cause or just patching symptoms?
→ before calling it done: what security hole or invariant did everyone overlook?
Opus leads · Sonnet builds · Haiku scouts · Fable challenges
JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split
the full tree
> Opus 5.5 coordinates the team and manages the shared task list at ~/.claude/tasks/
> Sonnet 5.5 runs builder teammates with direct peer-to-peer mailboxes at ~/.claude/teams/
> Haiku 5.5 runs fast scout teammates at 340 tok/s with zero coordinator lag
> Fable 5.1 stays on call as adversary to challenge contracts and audit diffs
> task dependencies unblock automatically as teammates mark work complete
> JEV routes micro-forks in milliseconds
paste the tree and this prompt into Claude Code ↓
"Rebuild my Claude Code setup around this tree:
1. Enable autonomous agent teams by setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json
2. Run the main lead session on Opus 5.5 to manage architecture and the shared task list
3. Spawn Sonnet 5.5 teammates for feature builds and Haiku 5.5 for 340 tok/s repo sweeps
4. Keep Fable 5.1 on call as adversary to audit peer-to-peer messages and test failures
5. Add one rule to CLAUDE.md: teammates must consult Fable before locking API contracts, after two test failures, and before final PR sign-off
6. Route fast deterministic forks through JEV, show all changes as diffs first. No edits until I say go"
↳ https://t.co/M2gSWKYxqH
save it, you'll use this every single day
i asked claude to cut an apple until it was a single f*cking atom
it took 104 cuts
this is the dumbest video i've ever made. it's also the most scientifically accurate thing on your timeline today
661 cubes. one bacterium. one virus. one strand of DNA. one atom. every size is real, the knife is the only fake shit in here
GPT 6 Astra can get worse when you give it more context
4 GitHub repos that stop a huge context window from turning into token gravity
trick: more tokens do not automatically mean more useful information
past a certain point the model starts searching through noise instead of reasoning over signal
here's the setup:
→ Repomix packs an entire repository into structured AI ready context
https://t.co/ARuX3AjAyH
→ Serena retrieves and edits code
semantically instead of dumping every file into the prompt
https://t.co/u2PyNyr4YM
→ Context7 injects current library documentation only when the agent actually needs it
https://t.co/A5pTpvvrEI
→ RAGFlow retrieves the relevant evidence before it reaches the model
https://t.co/02BhVcUWrG
the architecture is simple:
raw context → compress → retrieve → inject only what matters → reason → refresh
the goal is not the biggest context window
the goal is the smallest context that still contains the answer
paste this into Claude Code ↓
"Build a context engineering layer using Repomix Serena Context7 and RAGFlow
Use Repomix to create a structured representation of the repository
Use Serena for semantic code retrieval so unrelated files stay outside the active context
Use Context7 to fetch relevant library documentation only when required
Use RAGFlow to retrieve external knowledge before sending context to the model
Create a context budget for every request
Rank information by relevance
Remove duplicate stale and low value context
Show exactly what enters the final prompt before execution"
bookmark this before dumping another 100k tokens into your agent