KARASU
The crow
Scans the last 24-36 hours of AI news and picks the three strongest topics, each scored out of 25 and backed by at least two credible sources.
Watch a morning run
One real run, replayed in a minute
This is the run from 3 October 2026, step by step, with its real outputs: the topics KARASU scored, the captions Hisho wrote, the poster Horishi built and the mistakes Wakagashira caught and fixed before the post reached me.
KARASUTrend scan
Times are indicative; outputs, scores, fixes and numbers are the real ones from that run.
The problem
Posting well about AI, every single day
Vizuara teaches AI from first principles across several products. Posting about AI daily is hard to do well. The news moves fast, every claim has to be checked, each product has its own voice and look, and each platform has its own rules and limits.
Done by hand, that is a few hours of research, writing and design every morning, and quality drops on busy days. I wanted a system that did the whole job to a consistent standard and left only the final judgement, and the publishing, to a person.
How it works
One scheduled run, Monday to Saturday
At 9:00 AM a single scheduled run plays every agent in turn, following a written playbook. By the time I look, the post is finished and checked.
- KARASU scans the newsTop 3 topics, scored, with sources
- Kagaya picks the product and angleOr takes a product I chose in advance
- Hisho writes four captionsOne per platform, in the product's voice
- Horishi designs and renders the posterGIF + MP4, built from code
- Wakagashira checks and fixesAutomatic checks plus a visual review, up to 3 rounds
- The post lands on the deskReady to download
- NARADA reads the numbersLive counts from every post of the past week
- I publishAbout five minutes of my time
The gang
Eight members, one job each
Each agent has a name, a single responsibility, a written rulebook and hard rules it may not break. The names come from the roles in a Japanese crew: one boss, and specialists who each own one part of the work.
Kagaya
The strategist
Decides which product the story genuinely serves, which visual style it belongs to, and the one idea the post must teach.
Hisho
The writer
Writes four captions - LinkedIn, Instagram, X and YouTube - each following that product's own house style and the platform's limits.
Horishi
The designer
Writes a small program for each poster and renders it as a looping 1080 x 1080 GIF and MP4 in the product's visual language.
Wakagashira
The editor
Checks every caption and every frame, then fixes what fails itself and checks again, up to three rounds. Nothing unchecked gets through.
So-honbucho
The desk
Lays the finished post on a dashboard with download buttons, and keeps a desk for the live links once it is published.
NARADA
The journalist
Opens each live post in the browser every morning, reads the numbers and files a daily report, plus a weekly review on Mondays.
Shinobu
The human
Me. I review the finished post, publish it by hand and set the direction. Everything before that runs on its own.
The work
Posters the system designed
Each poster is a short program written for that day's idea, rendered as a seamless 1080 x 1080 loop. The complete poster is there from the first frame and only the meaningful parts move. These are real outputs from the morning runs.
The words
One idea, four platforms
For every post the writer produces a LinkedIn essay, a tighter Instagram version, a single sharp line for X and a YouTube description, each within that platform's limits and in the product's own voice. Below: the captions for "AutoCompact: coding agents learn when to compact their own context".
Deep into a long coding-agent run, the agent has read dozens of files, tried a couple of fixes, and its context window is almost full. Something has to go. Most agents handle that moment with a fixed rule: when the window fills, summarise everything and keep going. It sounds sensible. A new paper suggests it can quietly make the agent worse. AutoCompact (Zhang et al., submitted October 1) trains the agent to make three decisions itself: - When to compact, instead of waiting for overflow - What to keep: established findings, the relevant code and workspace state, and the actions still left - How to continue from that working-state summary The training runs in two stages: supervised fine-tuning on judge-corrected compaction decisions, then reinforcement learning rewarded only by task success. On Qwen3-Coder-30B-A3B-Instruct, SWE-bench Verified went from 30.4% to 39.6%. SWE-PolyBench Verified went from 19.5% to 24.5%. In the paper's 16K-window setting, the fixed-threshold baseline scored below the base model. It is one paper and one model family, so treat it as a strong signal, not a settled result. But the lesson is bigger than the number: context is not storage. It is a budget, and deciding what to forget is part of the agent's intelligence. That is the thinking behind Context Engineering from Scratch on Vizuara AI Pods, broken into five focused pods: Context Engineering - why the window is a budget, not a bucket Prompt Design - instructions that survive a long run RAG Systems - bringing the right facts back at the right time Memory Architectures - what to keep, compress, drop and recall Optimization & Evaluation - measuring whether your context choices actually help Skip these foundations and compaction looks like housekeeping. Learn them and you see it for what it is: a design decision that moved a benchmark by 9.2 points. What does your agent do when its context fills up? Tell us in the comments. Explore Context Engineering from Scratch on Vizuara AI Pods: https://pods.vizuara.ai/ #ContextEngineering #AIAgents #LLMs #CodingAgents #MachineLearning #ReinforcementLearning #AIEngineering #SWEbench #GenerativeAI
Your coding agent's context window is almost full. What should it forget? Most agents use a fixed rule: when the window fills, summarise everything and carry on. A new paper, AutoCompact (submitted October 1), lets the agent learn three things instead: - when to compact - what to keep: findings, relevant code and workspace state, actions still left - how to continue from that summary Trained with fine-tuning and then reinforcement learning on task success, it lifted SWE-bench Verified from 30.4% to 39.6% on Qwen3-Coder-30B-A3B. In the paper's 16K-window test, the fixed rule scored below the base model. One paper, one model family, so not settled yet. But the idea is worth keeping: context is a budget, and choosing what to forget is part of the intelligence. We teach this in Context Engineering from Scratch on Vizuara AI Pods, five pods from the context window to memory architectures and evaluation. Link in bio. #ContextEngineering #AIAgents #LLMs #CodingAgents #MachineLearning #AIEngineering #DeepLearning #ReinforcementLearning #GenerativeAI #LearnAI
Coding agents usually compact context on a fixed rule. AutoCompact lets the agent learn when to compact and what to keep: SWE-bench Verified 30.4% -> 39.6% on Qwen3-Coder-30B. Context is a budget, not storage. https://pods.vizuara.ai/ #AIAgents #ContextEngineering
AutoCompact: why coding agents should learn when to forget When a coding agent's context window fills up, most systems apply a fixed rule: summarise everything and keep going. AutoCompact, a paper submitted on October 1, trains the agent to decide for itself when to compact, what working state to keep, and how to continue. The working state it keeps has three parts: findings so far, the relevant code and workspace state, and the actions still left. Training uses supervised fine-tuning on judge-corrected decisions, then reinforcement learning rewarded by task success. Results on Qwen3-Coder-30B-A3B-Instruct: SWE-bench Verified 30.4% to 39.6%, SWE-PolyBench Verified 19.5% to 24.5%. One paper and one model family, so not yet replicated. Learn the foundations in Context Engineering from Scratch on Vizuara AI Pods: https://pods.vizuara.ai/ Paper: https://arxiv.org/abs/2610.02163
Every number in these captions traces back to the research paper the system cited. Captions are plain text, ready to paste.
The editor
It catches its own mistakes
Before anything reaches me, an automatic checker tests every caption and every frame, and an editor agent looks at the poster and reads the copy. When something fails, it fixes it and checks again. These are real catches from the morning runs.
- 3 Oct
AccuracyThe Instagram caption overstated a research paper's finding. Rewritten to say exactly what the paper reports. - 3 Oct
DesignFour layout problems on the poster: an accent line that read as an underline, two legend items almost touching, a 'full' bar that stopped short of full, and a missing colour key. All fixed and re-rendered. - 6 Oct
AccuracyA factual error: the caption mixed up experts per layer with experts across the whole model (10 of 512 in each layer, not 10 of 24,576). Corrected before it reached me. - 6 Oct
Platform fitThe LinkedIn opening ran past the point where LinkedIn cuts to 'see more'. Shortened so the hook is fully visible. - 7 Oct
Platform fitThe LinkedIn caption was 182 characters over the product's limit, then still 8 over. Trimmed in two rounds without losing the argument. - 7 Oct
AnimationThe poster visibly jumped each time the loop restarted. Re-timed twice until the loop was seamless. - 7 Oct
FormattingCaptions carried formatting asterisks that paste as raw symbols on social platforms. Removed, and a permanent rule added so it cannot happen again.
The numbers
A journalist that reads the results
Every morning NARADA opens each published post in the browser, reads the counts the page shows and tracks how they grow through the week. These are early readings from the first published posts (535 views across platforms); "-" means the platform does not show that number on the post.
| Post | Platform | Views | Likes | Comments |
|---|---|---|---|---|
| AutoCompact: coding agents learn when to compact... | - | 1 | 0 | |
| AutoCompact: coding agents learn when to compact... | - | 4 | 0 | |
| AutoCompact: coding agents learn when to compact... | X | 335 | 9 | 0 |
| AutoCompact: coding agents learn when to compact... | YouTube | - | 6 | - |
| Strata runs a 125B mixture-of-experts model on a... | - | 7 | 0 | |
| Strata runs a 125B mixture-of-experts model on a... | - | 3 | 0 | |
| Strata runs a 125B mixture-of-experts model on a... | X | 200 | 1 | 1 |
| Strata runs a 125B mixture-of-experts model on a... | YouTube | - | - | - |
Decisions
Why it is built this way
Posters from code, not prompts
Image generators misspell words and drift off-brand. Every poster is a short program, so the text, colours and logo are exact, and each design can be re-rendered or edited later.
An editor that fixes, not flags
An early version approved posts while only listing problems. I rebuilt quality control so it must apply each fix and pass the checks again, and every round is logged.
Evidence before opinion
Every topic needs two credible sources and every number must trace back to them. Results that have not been replicated are labelled as such.
No paid API keys in the loop
The whole chain runs without third-party AI keys, so there is nothing to leak and no surprise bill - in line with a credential-security standard I wrote for the team.
Each product keeps its own voice
Five products, five caption systems: hashtag rules, sign-offs, length limits and tone are written down and enforced automatically.
A human publishes
Nothing is posted automatically. The system ends at a finished, checked post; publishing stays a deliberate human decision.
- Claude Code (scheduled tasks)
- Claude in Chrome
- Python, Pillow, NumPy
- ffmpeg
- Node.js dashboard
- claude.ai artifacts
History
From a daily brief to a full studio
Version 0: a daily trend brief. The research agent ran every morning from July to September and wrote 60 briefs.
Designed the full gang of eight agents and the dashboard, and made it fully autonomous with the human only at the end.
Built the visual language: a campaign style studio and a library of 222 animated posters that became the designer's reference shelf.
Rebuilt the whole system as one scheduled run: code-built posters, four-platform captions, a self-fixing editor, browser-read analytics, a private live page, and five product voices.