Creao AI

Introducing CREAO: the Super Agent that delivers beyond the chat window. Describe what you need. CREAO builds it live. Save it as an agent. Run it on schedule. While you focus on what's next. Chat. Create. Run it all. #CreaoAI

AI agents are moving into production, and that means real exposure: unauthorized actions, data leakage, tampering, workflow hijacks. CREAO is partnering with @AgentGuard_AI to close that gap. CREAO builds and runs your agents. AgentGuard protects them at runtime, blocking unauthorized actions, stopping leaks, and catching tampering before it costs you. A place to run. A system watching how they run. That's production-grade agent infrastructure.

@PrajwalTomar_ GPT 5.6 and Fable 5 do not need to compete for the same job. The bigger win is tracking which one actually wins on a given task type over time, then routing there automatically instead of picking one model up front and living with it.

I still don't think people understand what just happened with GPT 5.6. Everyone is arguing about whether it beats Fable 5 and missing what actually died last week: the single-model era. The whole idea of one model doing all your work is over. Because for the first time, the math works. GPT 5.6 Sol ships code at half of Fable's price. Luna grinds through the small stuff for pennies. And Fable stays exactly where it belongs, planning the build and reviewing the diff like a staff engineer. One model was always a compromise. You were paying your smartest model's rate for work it was overqualified to do. Now you don't pick a model. You staff them: a manager, a worker, an intern. Same output, a fraction of the bill. My honest take: picking a favorite model is fandom. Staffing them is business. (full breakdown in the article below)

@0xCodila That token drop is really a caching problem in disguise. Anything that does not change between calls belongs in a fixed prefix, and once an agent's memory works that way by default, most of the cost problem disappears before you even touch routing.

Anthropic and Andrew Ng built an agent that uses 90% fewer tokens from scratch: they dropped the entire book of Frankenstein into a prompt - 108,000 tokens asked one question the input dropped from 108,000 tokens to 11 here's how: step 1 → put everything that never changes at the top - tools, then system, then docs step 2 → mark where the static part ends. everything above it gets cached step 3 → one stray space breaks it - and you pay full price again step 4 → the cache dies in 5 min - every read resets the clock step 5 → cached tokens don't count against your rate limits. free headroom most people never touch this - it pays for itself on day one watch & bookmark - this 1-hour brilliant course ↓

@supraEVM Interesting that the infrastructure bet starts at the versioning layer. The harder problem might be what the agent does between commits: context continuity, state across longer runs, memory that survives a restart.

GitHub's Former CEO Thinks AI Has Outgrown GitHub and Built a New Developer Platform Entire, the startup founded by former GitHub CEO Thomas Dohmke, has launched a distributed Git network built for AI coding agents. Instead of overloading GitHub, developers can mirror their repositories so AI agents clone and pull code from faster regional servers. The network is currently available in preview across the US, Europe, and Australia. It works alongside GitHub, not against it, but the long-term goal is to let developers host repositories directly on Entire. (Entire) @entirehq also saves AI coding history, including prompts, reasoning, and decisions. It introduces features like Entire Blame, Entire Review, and Semantic Search, making it easier to understand why code was written not just what changed. It already integrates with Claude Code, Codex, Cursor, GitHub Copilot, and other major AI coding tools. The company has raised $60M seed round at a $300M valuation, and plans to open-source its Git backend while expanding to native hosting and a fully decentralized network. As AI coding agents become more common, Entire believes traditional centralized Git platforms will become a bottleneck. Could distributed Git become the new standard for software development in the AI era?

@levantolabs When it gets something wrong and the agent still proceeds, what's the recovery path?

Meet the first AI to say ''I don't know''. We think agent mass adoption is blocked by three things: 1) weak security, 2) unreliability, and 3) how hard agents are to set up. Our first product, Sage, is focused on reliability. It's a "decision model" - a safer, faster, and cheaper way for machines to choose, act, and escalate to a human when confidence is low. You give it content (up to 32K tokens) and a list of questions (Sort, Yes/No, Choice, Tags, Scale), and Sage answers in 200ms - 9x faster than a traditional LLM - always with a confidence score attached. So yes… it's humble enough to say "I don't know." You can also turn on "grounding" to automatically run a web search and enrich the context. Under the hood: we took an open-weights LLM and fused on a classifier through post-training. It's great for agentic workflows, agentic guardrails, data pipelines, content moderation, operations, and risk & fraud. Why does this matter? Today's LLMs are great for chatbots, research, and creativity - but automation needs something much faster, with structured outputs, that isn't overconfident and is ready to admit when the signal is too weak. Sage preview is live. Excited to see your feedback. Levanto Labs is out of stealth today, founded by @marco_derossi and @bigironchris. We are hiring, reach out! Check the links in the post below 😊

@akshay_pachaar Picking the loop structure decides whether you're steering the agent step by step or handing that job to the system itself. That's the real design decision, and it's easy to miss when it all gets lumped under loop engineering as one thing.

the four types of agent loops. loop engineering keeps getting talked about as one thing. it's actually a choice between four structures, and each one fits a different kind of task. it means designing the system that steers the agent, instead of steering it yourself move by move. that system always answers two questions. what starts a run, and what decides the work is done. in a hand-run session you answer both yourself, every single time. each loop type moves more of that into the system. here's each type, what triggers it, and when to reach for it. 1) turn-based. triggered by a user prompt. the agent gathers context, acts, and checks its work inside a single turn, then a human reviews the output and writes the next prompt. use this when requirements are still forming and every output changes what you'd ask for next. 2) goal-based. triggered by a /goal command carrying success criteria and a budget, like "get the homepage Lighthouse score to 90, stop after 5 tries." when the agent tries to stop, an evaluator model checks whether the goal is met, and a no sends it back to work. use this when the outcome is measurable but the path there isn't worth your attention. 3) time-based. triggered by a clock. an interval fires, the agent runs a fixed prompt like "check the PR, fix CI," then waits for the next tick. /loop runs on your machine, /schedule moves it to the cloud so it survives a closed laptop. use this for recurring work where the task is known in advance and only the timing repeats. 4) proactive. triggered by an event or schedule with no human present. a routine watches a channel, and when something needs handling it spawns a workflow with a triage agent, a fix agent, and a reviewer that adversarially judges the work before the task closes. use this for standing responsibilities where you can't predict what will come in, only that something will. each type hands off one more job than the last. turn-based keeps both with the human, goal-based automates the checking, time-based automates the trigger, and proactive automates both while deciding the workflow shape at runtime. so the mapping question isn't which loop is most advanced. it's whether your task is exploratory, measurable, recurring, or standing. the more you hand off, the less you babysit. I wrote the full breakdown on loop engineering. the article is quoted below.

@cdngdev The extraction step is the sleeper move here. Turning an unstructured camera roll into a queryable wardrobe is where the real work happens before the fun part even starts.

If Claude's values shift by language, that's not a translation artifact, that's the model inheriting different value priors from whichever cultural context dominates the training data for that language. Worth knowing that deploying in a different language can shift more than tone.

In previous research, we found that Claude expresses over 3,000 values, like honesty and warmth. In new work, we asked how the values Claude expresses vary between Claude models and across languages. We analyzed 300K+ anonymized conversations to find out.anthropic.com/research/claud…

@nicochristie How does it deal with a spreadsheet that was built by five different people over the years, each with their own logic buried in it?

I challenged the MSFT Excel World Champion to a battle. AI beat humans at Chess, then Go. But those are games We built an agent to surpass humans on the most important app in the history of work Meet Shortcut: The Excel AI Agent Comment SATYA and I'll send you free credits

@bindureddy How does the loop tell the difference between genuine improvement and just cycling through six models without ever settling on one? That failure mode seems like it would be the hardest to catch from outside.

Currently working on AutoBots Completely autonomous cost optimized self-improving agents that do not require a human in the loop Simply set a bunch of goals and they will use a mixture of 6 LLMs to constantly improve and reach them

@hwchase17 The harness usually matters more than the model because it's the layer that decides what the agent can even see and do at each step, the model just decides what to do within whatever the harness exposes. Most teams spend all their effort on prompts and none on that boundary.

> For me, agentic performance on spreadsheet tasks is the whole of Shortcut existence. So I'm going to beat competitors by obsessing over the harness. custom harness is the only way you will beat the labs agentic experience heres how to build one: langchain.com/blog/how-to-bu…

@kevinnbass The half life on "best model" claims keeps shrinking faster than most roadmaps account for. Eighteen months from revolutionary to unusable is also the best argument yet for not marrying your product to a single model's assumptions. x.com/CreaoAI/status…

Just a reminder that deepseek v3 came out 18 months ago and was considered revolutionary at the time but is basically unusable today There was a fierce debate at the time about vibe coding and the argument was that LLMs can never create anything new because they are limited by their training data Those engineers were partly right and vibe coding was a real nightmare, I cannot believe what we used to put ourselves through with Sonnet 3.5, but we also knew real new things could be made and that the detractors were wrong I guess us non-engineers saw it most clearly (I would like to think so at least) because we were astonished at the new things we could do without the programming background, and we had nothing to lose and everything to gain in our enthusiasm Then came Opus 4.5 earlier this year which changed everything Suddenly real production code became possible with so much less friction and headache Everyone complained endlessly about models being nerfed or quantized but there was a steady march of progress from 4.5 to 4.8 Now a completely new crop of models is coming out that is not quite a 4.5 moment but something close We are not only getting more polish but things are becoming a little freaky, entire isolated domains become possible to combine with small teams or even just one person into new applications that would have required large infusions of time and capital in the past My wife back in 2021 told me about AI but I was in healthcare, I thought she was being a little nutty, and she talked like something from a science fiction film was coming and she got involved in it early on She never stops letting me know that she was right, and she was The next year is going to be wild, the world is truly going to change over the next few years, the scale of disruption will be a combination of astonishing and awesome, but also catastrophic There are many amazing things ahead and extraordinary challenges and opportunities We are living inside one of one of the biggest revolutions in human history

@haider1 This lands right next to Claude Cowork, similar pitch, similar timing. The real test is whether the plugin-based connections hold up as well as a deeper native integration would.

i really like openai's new chatgpt "Work" feature not everything involves programming, as sometimes you need help with project planning, content review, research, PDFs, or spreadsheets so i'd summarize this: need an answer? chat need to code? codex need to plan, research, or review? work

Agree completely, though the SEO and ads examples are the easy cases, the metric is unambiguous and hard to game. Curious how you'd design the loop for something like brand sentiment or support quality, where black and white doesn't really exist and gaming the metric is easy to do by accident

HOW TO USE AI LOOPS TO RUN YOUR BUSINESS 24/7 A lot has been written about loop engineering for building products. Almost nothing about using loops to run the business itself. That's the bigger idea. A loop is when you give an agent a goal, a way to check its own work, and permission to keep trying until it hits that goal. Build. Verify. Repeat. Stop when the condition is met. Here's what it looks like in practice: 1/SEO loop You're position 30 for a term you want. The loop runs once a month, makes changes, checks where you rank, and keeps pushing until you're on page one. This is running in production right now on Inbox Zero. 2/Ads loop You're spending $100 a day and losing money. The loop tests creative, checks profitability, kills what fails, and keeps going until the account is in the black. 3/Eval loop Your AI feature is only 88% accurate. The loop keeps adjusting the prompt and swapping the model until it passes 90%. 4/LLM visibility loop People search in ChatGPT now, not just Google. Same loop, new scoreboard. Are we the answer or not? The whole thing hinges on one thing: a metric that comes back black and white. Where do I rank? Did it hit profitability? Did the evals pass? Give an agent that scoreboard and it runs for months. Loops used to run for 30 minutes. These run for a year. Take a step, sleep, wake up next month, take another one. You're basically hiring an agency that never sleeps, gets paid in tokens instead of invoices, and undoes its own mistakes when the number goes down. Full episode on @startupideaspod watch

Any-any multimodal probably matters less commercially than it sounds because most real workflows only need multimodal at the input side, understanding an image or a screenshot, not generating one back out in the same turn. The output side gets solved by routing to a specialized image model instead.

One thing I am kind of surprised by is that full multi-modal (any-any) models have not become a bigger deal. It seems Google is the only Lab releasing these, OpenAI uses selective multimodal capabilities, Anthropic famously has no multimodal output & open weights models are mixed

@emollick Watching the cursor move under agent control is the moment it all clicks..the progress for gaming evals has been great

Computer use in Codex got very good on PC. Asking it to do something on your computer and having the cursor move under the control of a ghost is one of the things that makes you viscerally realize how much work can be done by a disembodied intelligence with a mouse & keyboard.

The sensitivity problem is really an argument for keeping the data out of the weights entirely. Once something is baked into a fine-tune it is genuinely hard to audit, redact, or prove it was never memorized, versus keeping it in a context layer you can inspect and revoke access to at any time.

The biggest challenge right now with the topic of every enterprise having their own model is that your most valuable information and insights are not only always changing, but they’re often your most sensitive information. Your most sensitive information can’t be packed into a model usually because it contains data that not everyone gets to have access to, and you can’t keep your security layer inside the model or an agent. I think there are going to be 100X more use cases for custom trained models, especially inside of domain-focused products, but training a model per enterprise is going to be a lot harder than it looks.

Cost per token dropping is only half of it though. The other half is that cheaper tokens make it economical to let an agent try something five times and pick the best result instead of getting it right once. That retry economics shift is honestly bigger than the raw cost curve for unlocking new use cases.

The best way you’re going to continue to get large scale agentic adoption is by continuing to bring down the cost of intelligence. More use-cases open up for AI every time you can have lower cost tokens (for the same or better level of capability). Almost all information work in the future will involve an agent somewhere in the workflow creating, processing, reviewing, or classifying data in some way. This will happen sooner *or* later depending on the cost of tokens of frontier models. Whether this happens from closed or open models is somewhat incidental, but the key is just that it happens. It’s great to see so much innovation and different approaches in AI right now as there are so many more use cases to power.

A 60x token bill increase is really a proxy for how much production surface area moved from human-run to agent-run in five months. The more interesting number is not the spend itself, it is what fraction of their internal tooling now runs unattended versus still needing someone to kick it off.

Just the CEO of ClickHouse casually saying their token bill is up 60x since February. Yes, 60x. His argument is simple: ClickHouse can’t build the best infrastructure for AI customers if it barely uses AI itself. So they’re burning all those tokens to learn by actually building with models, agents, inference pipelines, observability, and massive AI data workloads – to make ClickHouse the leading data platform for AI applications. Which I think they already are. Very special company.

The spiral into full test coverage over a single nit is a real pattern with Sol. It treats any deviation from expected state as equally urgent, regardless of actual blast radius. That is useful during a final review pass when you want maximum scrutiny, but expensive mid sprint when you are still exploring the solution space. Routing a faster, more decisive model for early iteration and bringing Sol in for the last pass tends to fix it without losing the rigor.

Some thoughts about GPT-5.6-Sol after ~30B tokens: Sol is the most OCD model I’ve used thus far. It very frequently gets one-shotted by random nits in the codebase and writes a bunch of tests to fix it. Even with fast mode, it’s incredibly slow to do this kind of iterative development, especially when builds take really long. This by itself is not a bad thing, but the worst part is that after 2 compactions, it’s chasing the nitpick / useless goals I never told it to accomplish, rather than the main task. This behavior is so bad, I thought I was messing something up and tried codex, pi, and opencode to figure out if it’s a harness issue, but there is no meaningful difference between the three, which leads me to believe this is a model problem. AI code has this weird delayed release effect. You’ll only notice slop code 2 dev cycles into a codebase when you spend more time fighting with the code and on refactors than on shipping features. It’s possible that sol is better than 5.5 a couple cycles in, but tbd. My file deletion experience has also been similar to others: this is a dangerous model to let loose without guardrails. For instance, when performing a routine container upgrade, it accidentally printed out an env secret, then panicked and rotated ALL secrets (this is internal so not public facing, which was also documented), and proceeded to break everything, spending an extra hour fixing everything and redeploying everything else to use the new secrets. It also gets rid of files it doesn’t like. I have no idea why this is, but I think something about the reward model rewarded bookkeeping. Writing is another problem. 5.6 has a huge context bleed effect. It does not know how to write documentation and starts putting the specs in the documentation. If I ask it to develop a user sandbox for isolation, and also ask it to write documentation, it starts talking about specs and sandboxes in user-facing docs, which makes no sense. Fable is somehow much, much smarter in this regard. Frontend design has also not gotten better. Fable is still one generation ahead here. Overall, as a huge 5.5 user, I am not convinced that sol is a meaningful upgrade. It’s possible my practices need to change, but unfortunately it feels like I’m spending longer fighting with 5.6 than I did with 5.5. It’s like the model is so SO smart, but so hard to work with, compared to fable and even grok4.5 surprisingly. It’s clearly intelligent, but also just doesn’t care about what I ask it to do? (Is this supposed to be AGI feels like?) I hope the codex team fixes what possibly is a bad harness setup, because the benchmark numbers show a very different story from what I’m seeing while using the model.