Issue #2 ·
Anthropic's equity fund, a breach at OpenCode, and agents that tune their own harness
Welcome back. Over the weekend the news was less about new models and more about how AI companies are run, secured and talked about. As always, every item below was checked against the original source, and the numbers are exactly as the source gives them.
Top stories
1 · Business
Anthropic commits 5% of its shares to a public benefit fund
On October 18 Anthropic announced the Anthropic Public Benefit Fund, a wholly-owned LLC. Subject to stockholder approval it starts with 5% of the company's shares, and the board can add another 2% for every 50% rise in the company's value above a set baseline. Together with the co-founders' pledge to donate 80% of their equity and an employee donation match, Anthropic says roughly 20% of its shares are now dedicated to the public good. The fund's equity will not be used for lobbying or political giving. Labor market impact is the example given, an area where Anthropic says it has already committed $200 million.
Why it matters: Equity is a decision you make early, not a gesture you add later. If you want part of your company to serve a purpose, write the rule down while the stake is still small.
Source: Anthropic announcement · @AnthropicAI
2 · Security
OpenCode reports a breach linked to a large-scale distillation attack
OpenCode said on October 17 that a breach the night before, related to a large-scale distillation attack, affected 1.4% of OpenCode Console users. Only users who had external model providers linked were affected, and all of them were emailed. OpenCode said the incident was still ongoing and that it was working with Anthropic, OpenAI and Google to fix it.
Why it matters: Any tool that holds your model keys becomes a target. Give every tool its own key, put a hard spending limit on each one, and rotate keys as soon as a vendor reports a problem.
Source: @opencode
3 · Agents
Google Research open-sources Harness², agents that improve their own setup
Harness² treats everything around the model as editable: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. For each new task it proposes edits, tests them on that same task and combines what worked, without changing the model weights, and it carries lessons forward to later tasks. The README reports that one step improved GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB's partial pass rate. It is Apache-2.0 and works with OpenCode and Codex executors.
Why it matters: Most of an agent's quality comes from the prompts, tools and memory you wrap around the model. That layer is yours to improve, and it stays with you when you switch models.
4 · Economy
Jack Clark's IMF lecture, and a GDP headline he corrected
Anthropic co-founder Jack Clark gave the IMF's Per Jacobsson Foundation lecture on October 17, titled “Inside the Event Horizon: Why Everyone Should Take AI Progress Seriously”. It covered faster scientific discovery, robotics, machine-driven businesses and AI systems that may contribute to their own development. A widely shared post claimed an internal Anthropic model projected a huge jump in US GDP. Clark replied that it was not an internal model but Anthropic's public economic scenario explorer, run on its “extreme” setting.
Why it matters: A forecast is only as good as the settings behind it. Before you put a big AI number in a pitch deck or a plan, find out where it came from and which scenario it describes.
Source: IMF event page · Jack Clark on X
5 · Work
Tim Sweeney: 70% of building Unreal Engine 1 was typing
Epic's CEO wrote that about 30% of his time on Unreal Engine 1 went to deep work on hard problems and about 70% to typing and refactoring large amounts of code. He said it always seemed logical that AI would take over that 70% once it could match human quality. What surprised him more is AI solving hard technical problems and working inside complex 3D scenes.
Why it matters: Look at your own week the same way. Split the repetitive 70% from the judgment 30%, hand the first part to tools, and keep your own time for the second.
Source: @TimSweeneyEpic
Open source & open weights
- •VoiceStudio bills itself as an open-source, fully local alternative to ElevenLabs: voice cloning, voice design, dubbing, dictation, transcription and audiobooks in 646 languages. It has 75,738 stars and is AGPL-3.0, so check that license before you build it into a product. GitHub
- •ai-job-search is a job application framework built on Claude Code that runs on your own machine: it rates postings, tailors CVs, writes cover letters and helps you prepare for interviews (MIT, 53,757 stars). GitHub
- •CLI-Anything, from the HKUDS lab, aims to give any software a command-line interface that agents can drive (Apache-2.0, 54,695 stars). GitHub
- •ai-memory adds long-term memory to coding agent CLIs and helps hand work off between agents from different vendors (MIT, 10,603 stars). GitHub
Quick hits
- •Dean Ball, now at OpenAI, says that after seeing what is visible from inside the company he puts distillation at about 50% of the reason Chinese models are close behind, up from the roughly 30% he would have guessed before. @deanwball
- •OpenCode's open-source coding agent is still on GitHub's daily trending list, with 1,414 new stars today and 216,748 in total. GitHub
- •Google's Android team publishes Android skills: SKILL.md instructions that help agents with Android tasks where its evaluations show models underperform. You install them with the Android CLI (Apache-2.0). GitHub
- •Cursor's plugin specification and official plugins repo is on GitHub's daily trending list, with 12,803 stars. GitHub
- •Hark launched Hark Pro, its assistant app, by giving away $1,000,000 of pizza across the US, ordered by typing “order a free pizza” in the chat. To reduce fraud, the promotion required users to connect an email account. @hark_labs
Repo of the issue
Orca runs Codex, Claude Code, OpenCode or Pi side by side, each in its own git worktree, so you can give one prompt to several agents, compare the results and merge the best one. A mobile companion app lets you follow and steer the agents from your phone. On October 19 it had 96,902 stars (1,146 of them added today), it was on GitHub's daily trending list, and it is MIT licensed.
For founders
- 1Treat every AI tool as a place where your keys can leak. One key per tool, a spending cap on each, and a plan to rotate them the same day.
- 2Put your effort into the harness, not the model. Prompts, tools, memory and checks are what make an agent good, and you keep them when you switch models.
- 3Check where a number comes from before you repeat it. The GDP headline from Jack Clark's lecture came from a public tool on its most extreme setting.
From Agentvendo
Launching something? Our $5 Website Video turns your landing page into a 15-second, 1080p 3D fly-over with music, in 16:9, 9:16 or 1:1. Paste your URL and the MP4 arrives by email within 24 hours.
Get my video for $5 →