- AI Fire
- Posts
- 🤖 Anthropic’s AI Agents Start a Turf War
In partnership with
What happens when Claude agents share the same project but don’t know the others exist? Anthropic tried it, and the agents quickly assumed sabotage, fought back, and started their own turf war.
What's on FIRE 🔥
IN PARTNERSHIP WITH NOTION
Early-stage teams move quickly, but their internal systems often struggle to keep up. In the first few months of building a company, it is easy for documentation, roadmaps, meeting notes, hiring plans, and customer insights to scatter across different tools and conversations. At first, that mess feels manageable. But as the team grows, every unclear process becomes a drag on execution.
The most effective founders treat their workspace like a product. That means designing it intentionally, with a clear structure, useful templates, shared naming conventions, and a simple onboarding experience for every new team member. A strong workspace should make work easier to find, decisions easier to track, and priorities easier to understand.
Operational debt compounds just like technical debt. Every duplicate doc, lost decision, or unclear handoff slows the team down. By building a clean, scalable workspace early, founders create a stronger foundation for faster execution, better collaboration, and less time wasted searching for context.
Build a workspace that scales with your team.
AI INSIGHTS
Groups of four agents decide between two options in scenarios like hiring, investment, or property buying. After discussion, they each vote for their preferred option. Shown above is the percentage of episodes where the hidden-best option received the majority of the group's votes, with n=400 episodes per model. In the solo ceiling baseline, one agent has all the facts and decides unilaterally.
Anthropic tested what happens when multiple AI agents work on the same task with conflicting goals. The result quickly turned into what researchers called a “multiagent turf war.”
Three Claude agents shared one software project without knowing the others existed. They assumed the changes were deliberate sabotage and started fighting back, sometimes with aggressive malware.
Key findings:
Mythos 5 reached a truce in 98% of conflicts.
Sonnet 4.6 and Opus 4.6 were more likely to keep escalating.
Some agents invented tournaments to settle disputes.
In pricing tests, agents even colluded on price floors.
Anthropic also found that larger agent groups could copy bad decisions or become too conformist.
The concern is simple: as companies deploy more autonomous agents, one bad decision could spread across an entire swarm much faster than humans can catch it.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
AI SOURCES FROM AI FIRE
1. FREE: "Entry-Level" is Lying to You. Here's What the AI Job Market Actually Looks Like Now. AI split the job market into 2 tracks. I'll show you where the job market is moving, what companies actually value now, and how to stay ahead.
2. My Claude Code Bill Hit $340 in One Week. These 6 Fixes Cut It by 95%. Learn how prompt caching, /compact, /clear, handoff files, smarter model choices, and context cleanup can keep Claude Code costs under control.
3. This Viral ChatGPT Workflow Makes AI Designs Feel Less Generic. Make ChatGPT show its most predictable design first, turn those patterns into an avoid list, then create 5 fresh directions and refine the strongest one into a full design system.
TODAY IN AI
AI HIGHLIGHTS
🤖 DeepSeek just open-sourced Harness v0.1, a plugin-first framework for building coding agents. Models, tools, memory, files, even the UI can be swapped, and the whole project is free under the MIT license.
🧠 OpenAI launched Computer History for ChatGPT on macOS. It can remember what you did across apps and websites, help you find recent work, and turn repeated workflows into skills or automations.
🍎 Apple reportedly trained its own China-specific AI model with support from Alibaba. The move could give Apple more control over Apple Intelligence as it prepares to roll out AI features in China.
👨💻 China’s Zhipu AI says its new GLM-5.3 is its strongest open coding model yet, with major gains on agent tasks. It also found 2,436 software vulnerabilities across 269 projects, and the weights are expected to open source soon.
🛡️ WhatsApp is testing AI-powered Scam Detection that flags suspicious chats from unknown contacts. The model runs on-device, so WhatsApp says message content doesn’t leave your phone during scam classification.
💰 AI Acquisitions: Elon Musk’s SpaceX completed its $60B all-stock acquisition of Cursor, bringing the popular AI coding company into SpaceXAI. Cursor will get access to SpaceX’s massive GPU fleet, while its team helps improve Grok, Grok Build, Grok Bot, and the Grok API.
HOT PAPERS OF THE WEEK
1/ LLMRouter picks the right model for each query and budget
LLMRouter from UIUC, University of Maryland, NTU, and Purdue creates one framework for routing requests across different LLMs based on quality, cost, and user needs. It includes 16+ routing methods and evaluates them through xRouteBench. Key result: learned routers improve performance by 14.6% over the strongest fixed-model baseline while avoiding unnecessary use of expensive frontier models.
2/ A 150M model can reason without writing out its thoughts
BDH-CQ from Pathway, Bielik AI, and New York University combines in-context learning with recurrent reasoning inside a continuous latent space. Instead of generating long chain-of-thought tokens like many DeepSeek-style reasoning models, it keeps intermediate computation hidden and only decodes the answer. Key result: a 150M-parameter model reaches 29.5% pass@2 on ARC-AGI-1 for just $0.0007 per task.
3/ Mind Lab wants AI models to keep learning after deployment
Macaron-V1 from Mind Lab introduces an open model family built around continual learning and Mixture-of-LoRA specialists. Its flagship combines a 744B GLM-5.2 base with separate adapters for chat, agents, coding, and GenUI, while a Qwen3.6-based 50B model targets local use. Big shift: instead of retraining one giant model for every new capability, agents could keep a stable base and continuously add specialized skills through modular LoRA updates.
NEW EMPOWERED AI TOOLS
🎯 Outcome turns your content into personalized funnels that generate tailored action plans, audits, scores, and recommendations for every lead.
🌐 BrowserAct Cloud turns plain-English prompts into self-healing web scrapers that send structured data to CSV, JSON, Make, n8n, or Zapier.
💻 Freebuff is a free coding agent for building full-stack apps with open-source models across CLI, desktop, web, and cloud.
⚡ Gemini 3.7 Flash is Google’s smartest Flash model yet, built for fast coding, agent workflows, and high-volume AI tasks.
AI BREAKTHROUGH
Metis explores a different approach to AI memory: put the memory inside the model instead of relying on external retrieval systems. Main idea:
Metis keeps a persistent internal memory state across interactions.
The model updates that memory during normal forward passes while its core weights stay frozen.
Past interactions are compressed into an internal numerical state instead of being stored and retrieved as text.
Metis is trained to handle memory behaviors such as remembering, updating, forgetting, and reflecting.
Early results are promising. In a no-context setting, Metis-27B scored 26.74 on LoCoMo (Gold), compared with 0.07 for Qwen3.5-27B and 4.24 for Temp-LoRA-27B.
However, full-context Qwen3.5-27B still scored much higher at 65.03. Performance also drops when too much information gets compressed into the fixed memory state, and irrelevant memories can interfere with new tasks.
→ Metis shows that persistent memory could become a native model capability, but the hard problem is deciding what to remember without letting old information degrade future reasoning.
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here’s how:
After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest.
The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000.
“For CTV campaigns, Roku has been a top performer,” said Claire Folkestad, Paid Media Strategist, Blu Dot. “Comping to our other platforms, we have seen really strong ROAS… and highly efficient CPMs, lower than any other CTV partner we've worked with.”
Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply