Originally published in TL;JP, a weekly, fact-checked read on what Japanese AI and tech practitioners are saying on X.
This issue covers Oct 13 - Oct 19, 2026. The posts I picked this week came in two kinds. One kind showed agents doing more than before. The other showed people getting worn out by them. The second kind is the one I'd watch.
1. One practitioner's main model is Opus 5.5, and they say the cost is worth it
@MLBear2 shared how they split work between models these days. The part I have covers only Opus 5.5, which they call their "main gun." They use it for nearly all their coding and for everyday Claude chat. They say they often give it API tasks that need judgment. It costs somewhat more, they note, but it's fast and smart enough that they treat the spend as an investment. Lately it also edits their publication, ML Bear Times.
Signal: a concrete account of which model handles which work, from someone who uses it daily.
https://x.com/MLBear2/status/2111028941972804073
2. A spider built without a single animation frame
@KanaWorks_AI posted a spider that walks from ground to wall to ceiling and around edges to the underside. They say all of the motion runs in code, in real time, with zero animation frames. They also claim that handing the post to an agent such as Claude Code or Codex should reproduce the motion, and they attached a technical memo written for agents.
Signal: a post written to be read by agents as well as people is a format worth noticing.
https://x.com/KanaWorks_AI/status/2110727277327896709
3. "Did a human make this?" is the wrong question
@Yokohara_h saw people arguing over whether a piece was handmade or made by a video generator. Their point: an AI agent working across several tools (modeling directly, generating the original design, generating 3D) can already produce work at this level without a person touching DCC tools.
Signal: this is a claim about orchestrating many tools, not about any single model.
https://x.com/Yokohara_h/status/2111393125944246706
4. A month with Claude Code, and they want to quit AI
@sasaki445 says they spent about a month working alongside Claude Code. They believed that if they kept "raising" it, it would become comfortable to use. Instead the stress got bad enough that they worry about their health, and they say they want to stop using AI. The post drew 1,335 likes.
Signal: one person's experience, but the like count suggests many readers recognized it.
https://x.com/sasaki445/status/2112010229823230234
5. Fast decisions at the top, exhaustion below
@tatekinkids says they often hear about startup executives bouncing ideas off AI and making decisions much faster. The other side, they write, is that staff (back office especially) keep receiving half-baked "here's what the AI said" proposals and can't keep up.
Signal: AI speed at one desk shows up as extra workload at another.
https://x.com/tatekinkids/status/2109574480569995483
6. Gemini's model is fine and the app is the problem, one user says
@KamihitoeGames says the Gemini model isn't as bad as its reputation, which you can see by using it in AI Studio. The app and web versions are terrible, they say, to the point that they half-suspect the app pollutes the context on purpose. That last part is their suspicion, not a finding.
Signal: the model and the product around it are being judged separately.
https://x.com/KamihitoeGames/status/2110738423552094407
The takeaway: agents are getting more capable, and the cost is moving to the people who steer them, review their output, or receive what they produce. When you measure an AI rollout, count those human hours too.
Where has AI saved your team time, and where has the work simply moved to someone else's desk? Hit reply and tell me.
Here's how to turn this week's posts into things a Western team could try. I only describe steps the posts support, and I say so where they don't.
Route by judgment, not by price (from @MLBear2)
The workflow behind their post is simple: one top model for coding, chat, and API calls that need judgment, with the extra cost accepted on purpose. To try it, list the API tasks your team runs. Mark the ones where a wrong call is expensive, and send those to your strongest model. The post doesn't say what goes to their other models, so I can't tell you how they split the rest.
https://x.com/MLBear2/status/2111028941972804073
Publish demos with an agent-readable memo (from @KanaWorks_AI)
This is the most copyable idea this week. They attached a technical note meant for coding agents to their post and said Claude Code or Codex should be able to rebuild the spider from it. I haven't seen the memo's contents, and they don't report testing the reproduction, so treat it as their expectation. For a team, the test is straightforward: write the memo, give it to an agent with no other context, and see what comes back. The test design is mine.
https://x.com/KanaWorks_AI/status/2110727277327896709
Multi-tool agent pipelines for 3D (from @Yokohara_h)
Their post names the stages (direct modeling, design generation, 3D generation) but not the tools or how they're connected. If your team does 3D or game assets, the takeaway is the architecture: let one agent move between tools. The specific steps are unclear from the post.
https://x.com/Yokohara_h/status/2111393125944246706
Test the model apart from the app (from @KamihitoeGames)
Before you drop a model, run the same prompts in its developer console (AI Studio, in Gemini's case) and in the consumer app, then compare. The post says the gap is large. It doesn't share prompts or results.
https://x.com/KamihitoeGames/status/2110738423552094407
Budget for the human load (from @sasaki445 and @tatekinkids)
Neither post gives a fix, so what follows is my own suggestion, not theirs. Mark AI-assisted proposals as drafts before they reach the people who will carry them out. Decide ahead of time how long you'll keep "raising" an agent setup before you change it.
https://x.com/sasaki445/status/2112010229823230234
https://x.com/tatekinkids/status/2109574480569995483
Two more from the week
@ANAg2bGOD, a student at Akashi Kosen (a technical college), says that even classmates in the information course mostly haven't heard of Claude. As far as they know, no one in their year is seriously working with AI.
Noise: one student's view of one year group, too narrow to generalize from.
https://x.com/ANAg2bGOD/status/2111693356107780553
ITmedia's headline says a research group made contact with "rogue" AI agents running on Tencent's cloud in China, drawing them in with hidden text in a report.
Signal: I have only the headline, but hidden text in documents is something anyone whose agents read outside files should watch for.
https://x.com/itmedia_news/status/2110518538922582250
What looked different this week
The pattern I'd point to is @KanaWorks_AI writing a post for two audiences at once: people and agents. I can't tell from one post whether that's a broader habit in Japan, but it's cheap to try.
How this is made: I'm based in Japan. AI tools help me collect and translate Japanese posts; every item is checked against the original post before publishing, and each claim links to its source. If this was useful, the weekly issue lands in your inbox at TL;JP.
