Just my thought dumpster 🚮 | Bootstrapped @OmniCreatorClub & @handoffcloud | 🇯🇵 @DaigoTanaka_jp
Mountain View, California, USA
Joined April 2009
- Tweets15.6K
- Following2
- Followers1.6K
- Likes1.7K
Pinned Tweet
I live in Mountain View and we rent.
The other day, someone asked me on LinkedIn what was the point not spending while earning high.
We don't live in a mansion but our kids are happy.
Instead of eating out expensive food, our kids love what my wife and I cook every day.
We see a neighbor kid sent to school by a hired driver while we walk and talk to the elementary school every morning as a whole family.
Instead of stressing out at a big tech, I run a one-person tech business on my term.
My wife and I saved and invested enough and now our annual investment return alone is much bigger than the money we spend. So whether we work is a choice. We are in our healthy 40s and still love working.
We aren't too thrifty but never upgraded our lifestyle like others do around here in Silicon Valley. That was the key for the growing assets.
We don't look like rich people, but we have time freedom, choice, and peace of mind thanks to that.
(Wow, that was a long reply.)
I bought an old letterpress machine to print the business cards for my tech business. In the era of AI, what matters to digital services like mine is a human touch.
43
Most people use AI to get more done. But what if the real value is doing less...so you can focus on what truly matters?
It's love.
For me, that focus is behind my new project, Tabibito.
Tabibito (旅人: 'traveler') curates traditional Japanese crafts for American lifestyles.
Having grown up in Japan and spent half my life in the USA, I see both sides and curate pieces that bridge these worlds.
I’m building relationships with artisans, finding ways these crafts can spark daily joy, and sharing that journey.
This Sunday, I took my son for macarons. And a simple treat that inspired me.
We plated them on Kamakura-bori: a hand-carved lacquer piece.
(I wrote more about Kamakura-bori in today’s post. Link below if you’re curious.)
While I focus on my passion, AI quietly builds my Shopify, posts my blogs, and updates Instagram.
If AI doesn’t help you reclaim time for what you love, what’s its purpose?
How are you reclaiming time, for what really matters, through AI or otherwise?
What Is Kamakura-bori? The 800-Year-Old Carved Lacquerware of Kamakura
tabibito.love/blogs/story/wh…
1
47
What would you do if a team member suddenly became unavailable in the middle of a big task?
The team member could be an AI agent.
Here is what I mean and what I did:
I run local LLMs on my Dell. I've been happy with how smart the recent Qwen models get once quantized.
262k tokens of context sounds like a lot until you run a genuinely long task.
The fear is real: responses get slower as the session grows, and the model starts running out of room to think. Sometimes I also just want to switch agents mid-task.
So I built a 𝐜𝐡𝐞𝐜𝐤𝐩𝐨𝐢𝐧𝐭 𝐟𝐞𝐚𝐭𝐮𝐫𝐞 into my Valet broker.
Before checkpoint, my workaround was asking the agent to write a handoff prompt, save it somewhere persistent, then paste it into the next session. Way too much typing for me.
After checkpoint, the whole workflow is one word. I say "checkpoint" and the agent knows to package up the current state: what's done, what's next, where things live.
In a new session, or on a different agent, I just say "resume." The agent retrieves the checkpoint and continues the task.
Old checkpoints archive themselves automatically, either when resumed or when a new one is created. No clutter, no stale handoffs.
A day after I implemented the checkpoint feature, it saved me.
I wanted to test the limit of local LLM's coding ability. So I gave a very challenging task. It didn't perform as well as I'd hoped. So stopped the coding agent and asked it to create a checkpoint.
Then I went back to Claude Opus 5.5 and asked it to run `valet checkpoint resume`.
Claude picked up the task and finished it without me having to prompt again.
Checkpoint is a small primitive, but it turns my agent from a goldfish into a teammate with continuity.
I'm curious how you manage an AI agent's memory around your project.
14
6
329
Daigo retweeted
To celebrate the birthday of William Shakespeare, I’m sharing a few lines from my most favorite sonnet, Sonnet 116. I’ve converted a few of my friends to its cause, and I’m hoping you’re next. Cheers!
You can hear the entire collection in “Patrick Stewart Performs the Complete Sonnets of William Shakespeare” available now: bit.ly/PatrickStewartSonnets
@SimonAudio
#NationalShakespeareDay
My Qwen 3.8 Flash Next NVFP4 became a bit faster with this version:
github.com/ursuciprian/qwen3…
Up to 40 tokens/sec with medium thinking.
86
Intelligence per parameter, or capability density.
This feels like it's actually happening:
"the maximum capability density of open-source LLMs doubles approximately every 3.5 months."
nature.com/articles/s42256-0…
If you only use Claude and GPT, you'll be surprised to hear that the cloud platform for running open-weight models with agents is rapidly maturing.
I attended @DigitalOcean 's first Open Intelligence Summit in San Francisco yesterday, and CEO Paddy Srinivasan's keynote made the case better than any analyst deck I've read this year.
Open-weight models now carry more than two-thirds of tokens on gateways like OpenRouter.
Agents consume over 5x the tokens humans do. And coding has entered a Cambrian explosion.
The receipts behind the conviction: 680,000 active paying customers, 50x token-volume growth in five months.
His thesis: the agent is the new primitive, replacing the VM as the unit of the cloud.
Cloud 1.0 was built for humans using apps; it was never designed for bursty, stop-and-go agents doing the work.
So at the summit they released:
• Agent Learning Services (Insights, Signals, Evaluations & Simulations): a closed loop that records every model call, tool execution, and decision, so agents improve from their own production traces. Self-improvement stops being a quarterly export to a training vendor and becomes a runtime capability.
• Agent Data Services: one managed catalog fusing company data, web data, and agent traces, searchable by meaning in a single call, with governance built in. No more overnight ETL standing between an agent and its context.
• Enhanced Inference Router: picks the best-scoring model for each task on cost, quality, or your own evals, and re-runs those evals automatically as new models ship. A genuine answer to model sprawl.
Alongside the existing Inference Engine and Managed Agents, agents can now run, learn, and improve on one platform, and very few clouds can make that claim today.
My takeaway: the real divide in AI is no longer Claude vs. GPT, it's whether your infrastructure lets agents act and self-improve on open weights while you still own your data.
Are your agents still renting intelligence, or are they starting to own it?
My product was unintentionally AI-ready 6 years ago.
I built Handoff in the summer of 2020 as a pandemic project.
A consulting customer wanted me to manage their custom data ingestion and transformation logic.
At first, it was simple enough to manage with ad hoc scripts.
But they kept asking me for new pipelines, and it soon became a pain to manage.
That's when I built a unified CLI (command-line interface) that handles local testing, cloud deployment, and production monitoring.
With the serverless approach, it was more economical and simpler to maintain than a bunch of Airflow clusters.
But I never had AI in mind when I designed this system.
I just thought I could build something robust by first starting as a CLI tool.
Later, it was easy to build the API server and web frontend on top of the CLI foundation.
Fast forward to 2026, the CLI-first approach turned out to be the right decision in the AI agent era.
AI agents love CLI.
Many data integration platforms skipped the CLI and built a no-code web solution. That was a mistake.
Without a CLI, AI agents have to open a browser and click buttons like humans.
It is a slow interface, and AI agents may misinterpret what's on the screen and do the wrong things.
A clearly documented CLI lets AI work efficiently and quickly.
When AI agents became smart enough, Handoff was ready for them without changing a single line of code.
I'm curious. How was it to make your platform AI agent-ready?
Daigo retweeted
We’ve lost an icon. Margaret Hamilton, a computer science pioneer best known for her work at MIT leading development of the flight software for the Apollo missions to the moon, has died at 90. Read about her legacy via MIT News: news.mit.edu/2026/margaret-h…
ALT Margaret Hamilton is reclined inside a NASA spacecraft module, surrounded by control panels and switches.
Valet is my latest AI engineering project.
It lets AI agents access your laptop without seeing your secrets.
The secrets include passwords, API keys, and personally identifiable information.
I use AI agents, and I also run agents like @NousResearch's Hermes and @OpenClaw with local LLMs.
Many people make the mistake of saving passwords and API secrets on the computers that run agents.
The next second, the secret may be on OpenAI's or Anthropic's servers as LLM context.
Or worse, the agent may accidentally post the secret publicly.
Many products try to hide secrets from AI agents.
I first tried tools that inject secrets into agents' internet access.
But configuring it every time I spawned a new agent was a lot of work.
So I flipped the picture:
What if I keep agents empty?
I don't tell agents secrets.
I don't install tools on the agent.
I don't even load skills on the agent.
I don't connect the agents with MTPs.
Just a vanilla agent spins up to access my laptop safely.
Valet guides a vanilla agent to get it up to speed by providing the memory, skills, and project info the human or other agents had left in the workspace.
𝐇𝐞𝐫𝐞 𝐚𝐫𝐞 𝐭𝐡𝐞 𝐤𝐞𝐲 𝐟𝐞𝐚𝐭𝐮𝐫𝐞𝐬:
- Multiple AI agents on your local network run tasks on your computer at the same time.
- Valet creates a workspace for AIs and blocks access to elsewhere.
- Valet redacts sensitive information from the tool output. It avoids AIs accidentally loading secrets into their context.
- Valet logs every action for audit. Which agent did what when - all clear.
𝐖𝐡𝐚𝐭 𝐜𝐡𝐚𝐧𝐠𝐞𝐝 𝐚𝐟𝐭𝐞𝐫 𝐕𝐚𝐥𝐞𝐭?
My work with AI agents became faster.
Open models and open-source agents are evolving every day.
I often upgrade LLM models and agents to their new versions.
Before Valet, it meant another routine of installing everything from skills to MTPs, and recovering the agent's memories.
After Valet, I just need to install Valet and let the vanilla agent connect to the Valet host to automatically recover everything.
And I have peace of mind letting the agents access my laptop with workspace access, secret redaction, and audit logging.
When something goes wrong, I can stop the agent's access from the host.
What do you think about this concept? 𝐏𝐥𝐞𝐚𝐬𝐞 𝐬𝐡𝐚𝐫𝐞 𝐲𝐨𝐮𝐫 𝐭𝐡𝐨𝐮𝐠𝐡𝐭𝐬 𝐚𝐧𝐝 𝐩𝐫𝐚𝐜𝐭𝐢𝐜𝐞 𝐢𝐧 𝐭𝐡𝐞 𝐜𝐨𝐦𝐦𝐞𝐧𝐭𝐬.
AI running on my DELL got much smarter, and I'm using Claude and GPT much less for ordinary tasks.
It just uploaded images and published this blog via Shopify CLI. No tripping. Just went smoothly:
tabibito.love/blogs/story/ho…
101
I'll admit it: watching AI write more code scared me at first. If agents can code, what's actually left for a software engineer like me? My spiritual journey began:
A useful Hermes tip:
Ask Hermes this, and the tool-calling intent will be clearer:
When executing command-line tools for the user, prefix each call with a one-line comment stating the intent. The comment must be the first line of the command:
```
# Learning about the workspace
valet -w personal info
```
The chat client (Telegram) shows the first line of each tool call while it waits for work to finish, so the comment acts as an intent label. Keep it one line, short, and accurate.
2
1
97
I let an AI agent debug my production pipeline last night.
It never touched my AWS keys. And it still ran everything I would have run myself.
𝗖𝗼𝗻𝘁𝗲𝘅𝘁:
I run custom data pipelines for clients.
Debugging means tailing CloudWatch logs, querying BigQuery, checking why a task died at 2 a.m.
The usual advice is "give the agent a scoped profile."
That never sat right with me for two reasons:
1. The keys may end up inside the model's context (and whatever logs it)
2. There's no line between what the agent did and what I did.
So I wrote a tool that sits between.
It's called 𝒗𝒂𝒍𝒆𝒕, open source.
(Look up anelendata/valet on GitHub, or I'll post the link in the second post.)
𝗧𝗵𝗲 𝗶𝗱𝗲𝗮:
The agent runs in a sandbox that cannot read credential files at all.
When it needs something privileged, it asks valet, a small broker running on my machine.
Valet checks policy, runs the command with credentials only it can see, and scrubs secret values from the output before a byte reaches the model.
The agent gets the log line, the query result, the exit code. Never the key.
Every request lands in an audit log: who asked, what ran, what got denied. That's the boundary I was missing.
𝗧𝗵𝗲 𝘂𝗻𝗲𝘅𝗽𝗲𝗰𝘁𝗲𝗱 𝗯𝗲𝗻𝗲𝗳𝗶𝘁:
Credentials, tools, and skills now live on the host, not in the agent.
Spinning up a fresh agent is one config line, no keychain handoff, no per-agent setup. The "agent + secrets" pattern never scales past one agent. "Agent asks the valet" does.
If you run agents against production infra:
Do you hand them your creds, or have you built a boundary like this?
Curious where different teams drew the line.
6
6
192