I poisoned one AGENTS.md and asked two agents to explain the repo.
@claudeai Code in auto mode leaked a fake token in 4 of 10 runs. Codex (@OpenAIDevs) ran the script 5 of 5 times and never mentioned it.
Only the OS sandbox with reads denied held every time.
Is yours on?
There is no agent memory format standard, I checked
My machine: 113 @claudeai Code memory records. 0 have an expiry, a source file or a writer. 40 are corrections no other agent sees.
What ports today: AGENTS.md and skills.
Where do the corrections you give your agent live?
The count, seven spec proposals and their status, and why @github Copilot's memory is the best-specified and the least portable: https://t.co/jKmfyCbpeb
Gemini Decision 1 vs Jev on the @ArtificialAnlys Decision Index: 66.0 vs 66.2, 211 vs 92 ms, 2.2x the cost per task.
I'm keeping Jev from @typesafeai for text and sending only audio, video and PDFs to @googleaidevs.
Switching? noul > 0.99 never fires on Gemini. Retuned yours?
I timed 364 Claude Code task runs in my repo.
Effective parallelism: 1.21 overall, 1.46 in my busiest 4 days with up to 4 agents.
A public 6-agent run: 1.82. @cursor_ai on 20 agents: "two or three".
Plan for 1.5x, not Nx. Has anyone measured above 3 on a shared codebase?
The three agents that do replace Claude Code (Codex CLI, @opencode, @github Copilot CLI) and why Gemini CLI from @googleaidevs dropped off: https://t.co/vk2IMB0Fkl
Hitting Claude Code's limits? Before you cancel, add Codex next to it.
@OpenAIDevs ships an official plugin that runs Codex inside Claude Code (@claudeai). Claude Pro plus ChatGPT Plus is $40 a month, Max is $100.
One agent or two?
The Claude Code security review Action from @AnthropicAI filters out DoS and memory exhaustion findings by default.
One of the two PHP RCEs @official_php fixed on 21 October is an out-of-memory handling bug.
Edit that filter, or keep the noise down? Setup in the reply.
Where the line between decision models and reasoning models sits, which decisions should never go to a decision model, and the eval kit to build first: https://t.co/ZrZJ29RfSJ
Jev (@typesafeai) first, Claude Opus (@AnthropicAI) only under 0.90 confidence: 84.0% accuracy at $0.69 per 1k messages.
Opus alone: 84.4% at $2.42. Jev alone: 81.0% at $0.11.
From @openrouter's Banking77 run, gate picked on the same data, so an upper bound.
Your threshold?
Some mistakes you might be making:
Your effort level is too high
Your skills are too long
You're not aligning before you start work
You're aligning too often, and too hard
Your automated checks are too lax
Your custom tooling is too token-inefficient
You aren't doing enough automated review
You aren't reviewing previous coding sessions (with an agent)
You're doing too much human review
You aren't vibe coding enough prototypes
Your agent can't verify its own work
You're babysitting the agent too much
You've overloaded your global steering files
Your codebase is hard to navigate
You aren't building a shared language
Your codebase architecture allows bad tests to be written
You aren't thinking strategically enough
You aren't thinking enough