William Cody Stanford

Director of AI | Techstars Alum | Neuroscience PhD | Learning maximalist

[William Cody Stanford @wcstanford]

https://t.co/AmokcZXyoE

[William Cody Stanford @wcstanford]

A completely buried aspect of this paper.

It's not about building a single harness.

They build K harnesses

Each to handle different clusters of tasks, and only evaluate edits on harness_i within cluster_i

The harness variant ensembling is where the gains are

[William Cody Stanford @wcstanford]

First refusal I've ever gotten from Anthropic (claude code).

Apparently generating descriptions of waste streams that involve anthrax is a no-no

I guess my eval set just won't contain BSL-1 agents...

[akira @realmcore_]

https://t.co/CudCkyWuRs

[Richie @richiemcilroy]

https://t.co/V5WdakwQbg

[William Cody Stanford @wcstanford]

@plugberryox @amytam01 I am on the exact opposite side of this.

It's never been easier to try things, fail, and use the learnings from that failure to refine your taste.

The people who leverage AI properly will become super-tasters.

[William Cody Stanford @wcstanford]

@Nikshep_09 @amytam01 In the long run, sure.

But often you have to try things that will fail in order to learn what ideas to say no to in the future.

[William Cody Stanford @wcstanford]

Taste is simultaneously the hardest skill to train while also being easier than it's ever been to develop.

[Amy Tam @amytam01]

https://t.co/tNsiY4fy87

[Amy Tam @amytam01]

https://t.co/tNsiY4fy87

[dotta 📎 @dotta]

We just open-sourced Paperclip: the orchestration layer for zero-human companies

It's everything you need to run an autonomous business: org charts, goal alignment, task ownership, budgets, agent templates

Just run `npx paperclipai onboard`

https://t.co/wuDdEmrSMx

More 👇

[William Cody Stanford @wcstanford]

If anyone's still looking for this, here's a plugin:

1. Auto updates docs based on code diffs
2. Enforces progressive disclosure of repo context
3. Plans as first class citizens
4. Migrates existing docs on init

PRs welcome

https://t.co/aPRJcu7LC0

[Suhail @Suhail]

It feels like someone should make a post-git-hook where it asks the AI model to look at the diff of what you changed for a merged PR and update the repo’s various readmes and other documentation to make it easier for an LLM to be able to write code and reference things faster rather than reading every single line of source code that might be relevant constantly. The agents need their own docs.

[William Cody Stanford @wcstanford]

@Suhail If anyone's still looking for this, here's a plugin:

1. Auto updates docs based on code diffs
2. Enforces progressive disclosure of repo context
3. Plans as first class citizens
4. Migrates existing docs to recommended architecture on init

PRs welcome

https://t.co/aPRJcu7LC0

[William Cody Stanford @wcstanford]

The insight is deceptively simple:

The bottleneck in AI development isn't AI capability.

It's human clarity.

Ouroboros fixes the human, not the machine.

[William Cody Stanford @wcstanford]

A redditor built a Claude Code harness that questions itself into completed products

It's called Ouroboros. 700 stars in a week

[William Cody Stanford @wcstanford]

Machine restarts? Doesn't matter.

The EventStore reconstructs the full lineage and picks up where it left off.

[William Cody Stanford @wcstanford]

RL your openclaw

Only 8 GPUs required

Still really cool work

[Ling Yang @LingYang_PU]

What if your AI agent got better just by talking to you?
Introducing OpenClaw-RL — a fully async RL framework that turns your everyday conversations into training signals. Your agent learns your habits, your workflows, your preferences. Privately. Continuously. #Clawdbot #openclaw

🔑 Two learning modes:
• Binary RL — likes/dislikes become rewards
• On-Policy Distillation — your textual feedback becomes token-level guidance

Self-hosted. Zero API keys. Your data never leaves your machine.
👉 https://t.co/ry18qekutm