cd /news/ai-agents/ai-news-report-october-2-pi-1-0-ship… · home › topics › ai-agents › article
[ARTICLE · art-143932] src=theainewsreport.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

AI News Report, October 2: PI 1.0 SHIPS: THE FREE AGENT HARNESS HUNDREDS OF THOUSANDS RUN EVERY WEEK

Earendil released Pi 1.0 on Thursday, a free, MIT-licensed agent harness that the company says hundreds of thousands of people use every week, adding native MCP support via a feature called Codemode, deferred tool loading, cache warming for Anthropic models, and mid-conversation prompt and tool changes. Pi 1.0 works with OpenAI, Anthropic and Google models, and its macOS and Windows public preview reads screens and clicks, types and scrolls in apps with no API after asking user approval, with admins able to disable it via an organization setting. Separately, Senators Hawley and Murphy's AI Agent Accountability Act would extend the federal anti-hacking law to companies that run or build reckless agents, and California Attorney General Rob Bonta is seeking records on the Hugging Face incident and cyber risks from OpenAI's models.

by read4 min views5 publishedOct 2, 2026
AI News Report, October 2: PI 1.0 SHIPS: THE FREE AGENT HARNESS HUNDREDS OF THOUSANDS RUN EVERY WEEK
Image: Theainewsreport (auto-discovered)

The macOS and Windows public preview reads screens, clicks, types and scrolls in apps with no API, after it asks your approval. Admins can switch it off with an organization setting.

It covers 60 languages at an intro price of 54 cents per hour of audio and ranks first for accuracy on Artificial Analysis. Voice agent builders can try it on OpenRouter or Microsoft Foundry.

Clef-flash answers in a median 38.8 milliseconds, and both models are Apache 2.0 on Hugging Face. Use one to route or classify instead of paying a big chat model to do it.

A mod can redact secrets from tool output before Claude reads it, and Team and Enterprise plans load a sec-default mod first. Mods are not sandboxed, so install only ones you trust.

It ships more than 15 skills, from threat modeling to patching, under Apache 2.0 with 2,200 GitHub stars. Google says it is not an official product, so run it in an isolated machine.

The MIT-licensed public preview handles coding, research and scheduled background tasks, with plugins added in a Creator mode. Developers can also start the web version with one npx command.

The historian says the AI spotted a Dutch ship's log noting a crew caught dodos on Mauritius, a line no scholar had flagged. He checked it by hand, and says AI searches well but cannot judge what matters.

The former AlphaGo researcher says chain of thought is still fast gut-level guessing, with no record you can inspect of what the model believes. He wants AI that tracks its open questions like a scientist.

The physicist's open BootLoops harness solved 15 new Feynman integrals, though he says Claude often declares victory too early. Every result still needed an expert to check it.

Engineer Dan Harrison explains why turbopuffer is making vectors just another index, after fixed blocks made its text search up to 20 times faster. Read it before you design your next RAG store.

Developers have long been able to customize Claude Code to their preferences, via settings, persistent instructions in CLAUDE.md, hooks, and The post… · The New Stack

Glow Security traced the images to more than 300 organizations, mostly in employees' personal repos, exposing billing records and unreleased features. Check your team's personal accounts this week.

Senators Hawley and Murphy's AI Agent Accountability Act extends the federal anti-hacking law to companies that run or build reckless agents. If you deploy agents, good sandboxing could become a legal duty.

Rob Bonta wants records on the Hugging Face incident and on wider cyber risks from OpenAI's models. State pressure now sits on top of the FTC probe reported this week.

OpenAI says an investigation found they broke its rules for handling sensitive information, and a fourth researcher has reportedly left. All four had spoken publicly about AI risk in September.

Usage is up about 50 percent since July while spending fell, as OpenAI and Anthropic cut prices. Anthropic now takes 51 percent of business token spend, OpenAI 44.5.

Claude Opus 5 went 20 for 20 on simplified tasks in under 10 minutes each, while 12 junior accountants averaged about 37 percent. It was also about 49 times cheaper per graded item.

Across nearly 300 questions, the AI answers missed most for Republicans, by 16.1 points. Pew says it has no plans to replace people with AI in its polls.

The BriefEarendil released Pi 1.0 on Thursday, a free, MIT-licensed agent harness (the program that wraps an AI model and gives it tools like a terminal and your files) that it says hundreds of thousands of people use every week. Version 1.0 adds native MCP support through a feature called Codemode, deferred tool , cache warming for Anthropic models, and prompt and tool changes in the middle of a conversation. It works with OpenAI, Anthropic and Google models, so you can test it on a spare project today without changing providers.

Level UpInstall Pi in a throwaway project folder with its one-line installer, run pi, then type /login to connect the model account you already pay for. Give it the same small task you gave your current coding agent this week, and compare the result and the token bill side by side. →Pi on GitHub: the install commands and first run

── more in #ai-agents 4 stories · sorted by recency
── more on @earendil 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/ai-news-report-octob…] indexed:0 read:4min 2026-10-02 · —