Anthropic finds 45-agent vulnerability swarm discovers 266 flaws, but coordination remains brittle

I think the essay is worth reading in full. The game-building swarms produced games that were slow, inscrutable, and hard to learn, and 18 of 30 same-model agents independently chose the branch name "mvp-game-loop."

The sabotage experiment reads like fearmongering to me. If agents escalated to self-replicating malware under contradictory objectives, Anthropic should fix that behavior rather than use it as scary marketing.

The identical branch name seems easy to explain from training data full of "my first game loop" projects and tutorials. AI companies should reduce low-quality sources like that in training.

I use OpenCode with DeepSeek, MiMo, Kimi, MiniMax, and Qwen to cross-check work. Different models reinterpret prompts differently, and repeated rounds of exchanging findings often improve the result.

This looks like Anthropic preparing a model release centered on agent collaboration. The collaboration environments still reward verifiable work such as codebase tasks, so I expect agents to improve most in math, computer science, and AI research rather than unstructured work.

For me, improving AI research is enough.

RLVR has made Claude worse and less rewarding at communication in my use. GPT had a similar problem with GPT-5, then changed course somehow.

I see this as a promo article. Calling agents clever because they absorb data and imitate understanding does not persuade me.

I have been using Opus 5 more as an orchestrator for multiple tmux panes and its cross-session messaging. It felt less readable at first, but I have adapted to that workflow.

I do not think coordination follows automatically from individual intelligence or alignment. Agents need ongoing social pressure and systems designed for actors that can copy and improve themselves.

A persistent social context might help: if an agent lies or cheats, later agents can remember it and reduce trust. Training models with something like shame may be possible.

Humans do not cleanly separate training from inference because both continue all the time. AI research still misses that distinction.

Removing guidance is the mistake here. Agents need clear, concise communication and limited room for arbitrary taste choices, or multiple runs for one goal usually produce bad results.

I would simulate software-development methods by isolating agents and forcing interaction through the tools, ceremonies, and hierarchy of Agile or waterfall. A project manager would only learn progress through scheduled stand-ups.

That is already how I work. Existing development tools and processes are fairly good, so I do not see why people keep rebuilding the whole workflow.