I build AI agents, ship browser LLMs, and let AI drive my kid's toy car. engineer turned product manager, still building software, ex-PayPal, startups. SF.
San Francisco, CA
Joined December 2009
- Tweets2.7K
- Following379
- Followers231
- Likes833
Sam Selvanathan retweeted
sharing my personal library on evals 1/n
i put together the highest quality blogs, podcasts, papers, and projects on evals. additions are welcome!
I set a goal in codex and the agent has been running for 9+ hours. Pretty impressed with the agent’s thinking and accomplishments! Thanks @gdb and team for this awesome tool!
With the ever growing and evolving AI tools that we have at our disposal, setting unreasonable targets and achieving those targets is not impossible anymore for anyone, especially for those building software products
Built a safer pattern for giving OpenClaw agents access to sensitive tools:
Don’t hand them full shell access.
Give them a narrow wrapper.
For my finance agent, I wrapped personalcapital2 / pc2 behind a Agent-only command:
fixed session path
creds outside workspace
allowlisted exec only
no bash/python/general VPS access
human-in-loop 2FA
The goal isn’t “trust the agent more.”
It’s: give the agent enough capability to help, while making the unsafe path unavailable by design.
Giving an AI agent email access sounds scary if the boundary is just “the prompt says don’t do bad things.”
I’m setting up a personal research agent with Gmail access, but the real design is the permission envelope: read-only OAuth, allowlisted tools, isolated workspace, and human approvals.
The goal is not “AI has my inbox.”
The goal is narrower: the agent can search and read relevant email so it can help surface important threads, summarize context, and track follow-ups without getting send/delete/archive powers.
The easiest way to create a new AI agent is to copy the main assistant and change the name.
That is also the fastest way to create a mess.
Specialized agents should not inherit broad context, broad tools, broad memory, and broad permissions by default.
Sam Selvanathan retweeted
More amazing tools from @mvanhorn. Been using these for a month and they are amazing.
Introducing the Printing Press, a CLI-factory and a CLI-library. Built with @trevin. 🏭🖨📚
Most APIs suck for agents. Most MCPs suck for agents. Most official CLIs suck for agents. They waste tokens and time. @steipete started making his own because of this.
📚 A Library of agent-native CLIs you install today (Linear, ESPN, Flight GOAT (Google Flights + Kayak nonstop), Contact Goat (LinkedIn + Happenstance + Deepline more) +30+ more)
🏭 A factory that prints new ones for any service - just type /printing-press <product name>
CLIs are fast, local, SQLite-backed. Work in Claude Code, Codex, OpenClaw, Hermes.
🌐 printingpress.dev
The core idea:
Don’t give the agent the secret. Give the agent a capability.
A repo-scoped PAT helps, but it is not enough if the agent also has arbitrary shell access.
teapawt.org/samselvanathan_/status…
I set up backups for my AI assistant workspace, and it turned into a useful disaster recovery pattern.
Two layers:
GitHub for working files.
Restic + Google Drive for full recovery.
Simple, boring, recoverable.
What I like about this setup:
boring cron jobs
private GitHub history
encrypted off-machine backup
clear retention
failure alerts
no heroic manual process
If it breaks, I want to know.
If the machine dies, I want a path back.