I like cats. I like jogging.
Singapore
Joined January 2026
- Tweets146
- Following84
- Followers9
- Likes177
Hanami retweeted
Check out anywhere with Shop Pay
Today, we’re expanding our partnership with @Shopify beyond finding, buying and tracking orders from Shopify stores. Now, your Instinct can use Shop Pay at any online store.
The card and address you've already saved in Shop Pay now work across stores that aren’t on Shopify, too. This makes it easy for your Instinct to make secure purchases on your behalf.
You approve every purchase in Shop Pay, and your card details never leave Shop Pay.
A small Chrome extension: labels every tweet by topic and folds those you don't read: ads, spam, repeats, politics, crypto, for free. stopslop.cc
Hanami retweeted
Nanbeige just dropped a 33B open agentic coding model worth watching
it is Apache 2.0, dense, and carries a 512K context. On their card it posts strong numbers against other ~30B models on terminal and software-agent tasks (Terminal-Bench 2.1 at 80.9, DeepSWE at 65.5). It trails slightly on SWE-Bench Pro
Practical notes from the card:
- BF16 weights are ~66 GB
- Recommended inference is multi-GPU (tensor parallel 4) with vLLM or SGLang
- Tool calling is supported (XML preferred)
- Thinking mode can be toggled; they recommend keeping reasoning for agent work
another open coding agent to put on the watch list
especially if you want something you can run and fine-tune yourself
Hanami retweeted
Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.
This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.
→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.
→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.
The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.
Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.
As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.
claude.com/resources/article…
Hanami retweeted
Local machines had become a bottleneck for agentic mobile development at @Shopify, so we gave agents their own remote dev environments. They now author 61% of our merged PRs 🧵
Hanami retweeted
Security called me at 6:12 AM today.
"Patch every PHP server immediately."
I asked what happened.
"Two critical vulnerabilities. Remote code execution. No authentication required."
Cool.
How many PHP servers do we have?
According to inventory: 12
According to reality:
Nobody knows.
We found one running WordPress.
Three running old internal tools.
One under someone's desk.
Two inside vendor appliances.
A forgotten employee directory from 2016.
And a server named:
PHP-OLD-DO-NOT-TOUCH
Naturally, that one runs payroll exports.
Then we found a PHP 5.6 box.
I asked who owned it.
The documentation said:
BRIAN.
Brian left in 2019.
We called Brian.
He said:
"Oh wow, that's still running?"
Excellent.
By 8 AM, leadership joined the incident call.
CEO:
"Can hackers actually get in?"
Security:
"Potentially yes."
CEO:
"Without a password?"
Security:
"Yes."
CEO:
"Then why do WE need passwords?"
Not the lesson.
Finance asked if we could wait until the weekend to patch because downtime affects revenue.
Security explained that the alternative was letting strangers execute code on our servers.
Finance asked for a cost comparison.
At 9:17, someone suggested temporarily taking the vulnerable websites offline.
Marketing refused.
"Our campaign is live."
So for the next hour, the company officially decided maintaining a landing page for 20% off webinar registration was more important than preventing remote server takeover.
We finally patched everything we could find.
Then our scanner discovered another PHP server.
Hostname:
TEMP-PHP-2014
Uptime:
4,512 days.
Nobody knows what it does.
Nobody owns it.
Nobody can shut it down.
But every time we block network access, the warehouse stops printing shipping labels.
Security has classified it as critical infrastructure.
It is running PHP 5.4.
The vulnerability has existed longer than half our employees.
The emergency patch took four hours.
Finding out what PHP actually runs took eleven.
The internet isn't held together with code.
It's held together with servers everyone is afraid to turn off.
Hanami retweeted
Clean handoffs don't just save tokens, they let you run most steps on a cheaper model without giving up performance.
When an agent only gets the spec (the plan for what to build) or the diff (just the lines that changed), its job gets much smaller than reading the whole history, and smaller jobs often don't need the biggest model to do them well.
Here's how I would apply that to the four steps in the post:
Planning is where I would keep the strongest model, because a vague spec makes every later step worse.
Coding gets the requirements and the relevant code, so a mid-tier model is often enough for routine changes.
QA gets the changes and the test criteria, which turns it into a focused check. For example, "does each acceptance criterion have a test?" is worth trying on a cheaper, faster model.
The fix-and-rerun loop needs a limit. Without one, the review agent and the fix step can go back and forth on the same failure and burn more tokens than the bloated context did. For example, allow two attempts, and if the same check fails twice, send the diff and the failing check to a developer.
At work, we pick the model per task based on performance and cost, and scoped handoffs make that easier because each step becomes a task you can test on its own.
Hot take: Multi-agent systems have a serious context problem. They're burning tokens for nothing.
Everyone is racing to build AI software factories with more agents: planning agents, coding agents, QA agents, PR review agents.
Sounds impressive. But adding more agents doesn't automatically make the system more efficient.
Here's the problem.
Too often, every agent gets a massive amount of context from the previous step. Full conversations, tool outputs, code history, reasoning, and information that has nothing to do with its actual task.
We're paying models to repeatedly process information that another agent has already processed.
That's not intelligent orchestration. It's expensive.
I think the better approach is much simpler.
> Clean the context at every handoff.
> Structure the output.
> Give the next node exactly what it needs to complete its task.
1. The planning agent produces a spec.
2. The coding agent gets the relevant requirements and codebase context.
3. The QA agent gets the changes and test criteria.
4. The PR review agent identifies actual issues, and another execution step fixes them and reruns the checks.
No need to carry the entire conversation through every node.
And no need to involve a developer every time an agent finds something wrong.
For low- and medium-complexity tasks, the system should be as autonomous as possible.
Engineers should focus on architecture, new domains, and complex product decisions, not babysitting agents through routine fixes.
The real bottleneck in AI-native software development is no longer generating code. It's orchestrating the entire workflow reliably and efficiently.
The goal isn't to simply have more agents. It's to get more completed work out of fewer tokens, with less human intervention.
That's where context engineering becomes more important than agent count.
Hanami retweeted
A few thoughts about the chief of staff pattern that I haven't seen before, but would love to learn more about:
- 90% of the time, all my agent needs between turns is a 👍 keep going buddy
- the next 9% is multi choice, often with an obvious best choice
- only the last 1% reauires ground up judgement from me
I've designed my Chief of Staff agent to primarily act as a judge BETWEEN each turn depending, deciding which of those 3 buckets is appropriate.
Bucket 1 keeps work moving.
Bucket 2 makes it easy for me to approve (and even automate some of those approvals).
Bucket 3 get escalated to me quickly.
My whole @jfdibot was fundamentally built around agent turns as a building block, which in hindsight turned me into the bottleneck to decide what happens between every turn.
This pattern feels like a magic trick in more ways than I can describe.
Also, as I've gotten better at working with agents (and thanks to lessons I learned working with @mattpocockuk), I've learned to front load the decision work and build in verification layers.
I just don't see this being done often outside of programming work, so I'm genuinely interested in more examples!
Hanami retweeted
Enterprise AI right now is kinda like Bush era foreign policy:
1. Go into a foreign country / company
2. Blow things up, depose current leadership
3. Enact regime stewarded by your experts
4. Have no exit strategy. Oops now you're running their shit or else it all collapses
Hanami retweeted
Coinbase is saving at least $10M a year by doing this.
Classification is easier than generation. Enterprises easily receive millions of queries per day. It will never be cost effective to use frontier LLMs to classify every single query.
Quick math on Coinbase fraud: ~7K tokens per check. That's ~3.5¢ per call on Opus, ~2¢ on Sonnet, ~0.05¢ on a self-hosted 9B. This means $4-16M a year running frontier on every transaction vs ~$80-200K for the 9B. A 9B model in fp16 is 18GB, fits on a single GPU for ~$1-2/hr.
Even if your GPUs sit at 10% utilization, a fine tuned 9B is still 3-7x cheaper per check than Sonnet. The real comparison is against small hosted models, and there quality and control are the deciding factor instead of cost.
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Hanami retweeted
One of the biggest bottlenecks in scaling AI is that many business systems, especially older ones, don't have APIs.
One underrated effect of Computer Use getting much better is that, all of a sudden, those systems become programmatically accessible, since agents can access them through the same interfaces humans use!
Hanami retweeted
our heroic members of technical staff
their brutish distillers
our glorious cyber capabilities
their wicked cyber risk
our noble human data pipelines
their backwards distillation attacks
Over the last few months, we’ve made significant progress in reducing the effectiveness of distillation attacks against Claude.
In this post, we explain how: anthropic.com/research/disti…
Hanami retweeted
Let me teach you what an AI harness is in 90 seconds, because once you get this, every AI product on the market starts making sense.
The model is the part everyone talks about. GPT, Claude, Gemini. Those are just brains in a jar.
They produce text and nothing else.
The harness is the invisible software wrapped around the brain that turns its text into real action.
Here's the loop that runs every single time.
> Step one.
The harness loads your goal, the rules, the history, and any previous results into one big pile of context.
> Step two.
It hands all of that to the model and asks "what's the next move?"
> Step three.
The model either answers you directly or asks for a tool. Read a file. Run a test. Search the web.
> Step four.
The harness checks the permission, runs the tool, and feeds the result back as the next input.
Then it loops. Over and over. Until the job is done or it hits a time, cost, or step limit you set.
That's it. That's the whole magic.
Every autonomous AI agent you've ever used is just this loop with a bigger brain and better tools strapped on.
Now you know what's actually under the hood.
Hanami retweeted
great read.
Personal AI agents (e.g. shopping agents) that research every purchase super exhaustively could make product quality matter more than distribution.
Agents that never forget to cancel would end subscription revenue built on customer inattention.
The asymmetry between a company and its customers comes from scale (i.e. a compute asymmetry), because the firm amortizes its analysis over millions of purchases while the customer pays the full cognitive cost of each one.
Coupons work only while search is costly, so agents that make search free either hand the discount to everyone or kill it, and the same logic drains the inattention revenue subscription businesses book as margin.
Assorted things I've been mulling over regarding agents in the economy:
• Companies are in some sense the original superintelligence. It's often the case that they can use their amortized cognitive surplus to advance their interests against those of certain consumers: relative to the optimization capacity of companies, consumers can be inattentive, underinformed, irrational, etc. (Think, say, of some of the classic insurance upsells at the end of a booking process, which is typically priced substantially above what a regular policy would cost.) Herbert Simon introduced the idea of bounded rationality, and firms often end up benefiting from it. Thanks to personal agents, we will probably get somewhat more rational in how we engage with companies.
• How exactly will outcomes change? Well, certain kinds of price discrimination will work less well (how do coupons work if every agent hunts fastidiously for them?), and certain kinds of customer cross-subsidy will change (today, people who forget to cancel subscriptions subsidize those who don't). Taken across the economy as a whole, I'm not sure what the incidence of all of this will be.
• In corporate strategy, people have long debated the relative importance of distribution and product quality. Perhaps you've made something better, but does anyone know about it? Won't people just continue to buy the ACME Corp Mousetraps? Personal agents will, I think, act as a kind of structural subsidy for product quality. How would markets look different if individuals always spent at least 10 hours researching their purchases? On the margin, focusing on making a product better is, I think, going to become a more effective strategy for companies. This is good!
• It true, a corollary is that information about product quality will become more important and impactful. Today, most people have very limited ability to aggregate and integrate such data, but agents will have insatiable thirst for such signals. (It's often said that data is becoming more valuable because it helps in the training of models. That's obviously true, but I think there's a secondary way in which it's more valuable simply as a complement to cheaper intelligence.)
• A big question is whether agents will increase superstar effects (rational agents all settle on the same product) or increase dispersion (because everyone has slightly different tastes and preferences, which agents are good at eliciting, finally overcoming the artificial clustering we've been subjected to).
• Many internet companies have monetized demand routing, which is in turn often predicated on short attention spans. "I'll click on the first thing I see." This has led to some amount of inefficiency, as companies try to propitiate capricious ranking algorithms and incur various placement fees. It's not all deadweight loss (willingness to pay is itself a kind of signal, which is valuable to buyers), but it's not great. How will the nature and value of this routing change in a world of personal agents? I don't know the answer, but I think that the effects might be pretty big.
• Much of what I'm writing hinges on the assumption that personal agents will be on the consumer's side. I suspect that that will be the winning strategy, but whether it's true, and what it means to be true, involves a lot of other questions.
Taken as a whole, the market is a massively parallelized, decentralized, and highly imperfect value attribution machine, and, taking stock of all of the effects, it seems to me that the quality point may well be the biggest: that agents will change the nature of the backpropagation such that the rewards for better products increase. Maybe this is too optimistic, but I suspect (and hope) that companies that make things that they're proud of will feel that the universe is a little more partisan in their favor.
Hanami retweeted
Most measures of AI progress closely fit an exponential over the 2023-2026 period. Meanwhile AI capex over the same period has been slightly super-exponential (i.e. the rate of growth itself is increasing).
So if you model AI as a system that takes as input external investment and outputs AI progress, that system has a slightly sub-linear response.
But super-exponential injection of *external* resources is not sustainable in the long-term. To achieve sustained exponential progress, the AI industry will need to achieve matching exponential profits -- it will need to close the resource loop.