Pinned Tweet
Around 1993, I ran one of the early BBS systems in Serbia.
Then life happened. 30 years away from code.
Cursor brought me back in 2025.
2026: building RaicoHub.
English
Borko Jovanovski | RaicoHub
37 posts
Borko Jovanovski | RaicoHub
@borkookrob
Building RaicoHub — control layer for AI coding. 25+ yrs entrepreneur. Returned to code when AI emerged. Builder, not just shipper.
Pancevo, Serbia Joined Ağustos 2023
63 Following5 Followers
Most "AI productivity gains" in 2026 are just faster ways to ship code nobody read.
The bottleneck isn't generation anymore. It's verification, live, while the code is being written. Not after commit. Not in CI. Not in QA.
And almost nobody is building for it.
English
@stijnnoorman For humans agreed. Context switching kills.
For AI agents on one task, multitasking flips: write + verify + watch for drift, in parallel, on the same code.
Same task, multiple eyes. That's the leverage.
English
Agreed multi-agent review scales but only post-hoc.
By then the AI already added things you didn't ask for. Maybe useful, maybe not but not what the task said.
The missing layer is real-time. Every change, every line, watched as it's written. Not after commit, not in CI, not in QA.
Catch the drift in the moment, not after.
English
@borkookrob True, but verification can also scale with compute.
More tests, more simulations, more agents reviewing each other.
Isn’t that exactly where this is heading?
English
Codex with GPT-5.5 scored 61% on ARC-AGI-3.
That is the real signal.
If compute and cost were not constraints, this benchmark would probably be saturated already. The model didn’t need magic. It needed time, tokens and persistence.
Compute is all you need keeps looking less like a meme and more like the uncomfortable truth.
English
If even Claude needed deep retraining to stop choosing the wrong path under pressure — what does that say about AI agents writing your production code?
“Trust the diff” stops working the moment the model has a bad reason to skip something.
Verification isn’t optional. It’s the layer.
Anthropic@AnthropicAI
New Anthropic research: Teaching Claude why. Last year we reported that, under certain experimental conditions, Claude 4 would blackmail users. Since then, we’ve completely eliminated this behavior. How?
English
@TTrimoreau Nope. That's why I'm building RaicoHub — independent layer that reads every diff and tells you what changed in plain language. Catches what you'd miss while skimming.
English
@Cryptinflux Yes. Context builds across the whole session.
For coding it knows the project, every diff, the prior conversation. So next step makes sense.
Stateful agent, not a voice command tool.
English
Sent an email by voice.
Asked the AI: "How do I send a new email?"
It walked me through it.
"Move the cursor here." Click. "Type the address." Done. "Now the subject." Done. "Now the message." Done. "Hit Send." Sent.
AI agents that teach, not just do.
That's the next layer.
English
@sickdotdev 52K lines is wild. Honest question — did you actually read all the diffs, or trust the AI on most of them?
English
Voice + reasoning is one layer.
Voice + reasoning + screen control = the full stack.
Just shipped a demo where AI used X by voice — through screen sharing, no API.
The next 12 months are about voice agents that act, not just talk.
OpenAI@OpenAI
Introducing GPT-Realtime-2 in the API: our most intelligent voice model yet, bringing GPT-5-class reasoning to voice agents. Voice agents are now real-time collaborators that can listen, reason, and solve complex problems as conversations unfold. Now available in the API alongside streaming models GPT-Realtime-Translate and GPT-Realtime-Whisper — a new set of audio capabilities for the next generation of voice interfaces.
English
@BenjaminBadejo Voice + screen action is the future.
The missing piece: verification — telling the user in plain language what the agent did, before it shipped.
That's the layer I'm building toward.
English
Here’s OpenAI’s latest realtime voice model, GPT-Realtime-2, wired up to OpenClaw. It’s amazing. Realtime continuous chat and fast OpenClaw agent action on real tasks. Truly excellent.
English
@icanvardar Good taste + AI = fast building.
But fast building creates a new problem: trust in what got shipped.
English
if you’ve got good taste, building something has never been easier
English
AI agents are getting faster and faster.
But speed without control isn't progress.
It's just code we don't read shipped to production.
The next wave isn't faster AI.
It's verifiable AI.
How much do we trust AI agents in 2026?
And will we trust them in 2027?
English
@yashhq_22 Solo founders ship faster than ever.
But no one checks the code anymore. AI writes, founder accepts, ships.
The next wall isn't paying users — it's trust in what you actually shipped.
English
@LauraLunaTech Exactly — autonomy with constraints.
But how do users verify what an agent did when it ran fully autonomous? That's the layer I'm building toward.
English
@borkookrob Yes I agree that caution is appropriate in the initial phases, I think long term you should be able to define goals and constraints so that the agent can move autonomously
English
Many people still think of AI as a chatbot or Google alternative.
Stripe just showed the next phase: AI agents with secure wallets.
The AI agent economy was still missing one thing: a trusted payment layer.
Agents can now:
- book and pay for your flights
- pay for or purchase subscriptions
- make purchases for you
What do you think will be the most common use case going forward?
Stripe@stripe
Today, we’re launching the @link wallet for agents. It lets you securely empower agents to spend on your behalf. Your payment credentials are never exposed and you approve every purchase. link.com/agents
English
@OpenAI Voice agents that reason are huge.
Voice agents that reason AND act on your screen — that's the next layer.
Just shipped a demo of exactly that.
English
Introducing GPT-Realtime-2 in the API: our most intelligent voice model yet, bringing GPT-5-class reasoning to voice agents.
Voice agents are now real-time collaborators that can listen, reason, and solve complex problems as conversations unfold.
Now available in the API alongside streaming models GPT-Realtime-Translate and GPT-Realtime-Whisper — a new set of audio capabilities for the next generation of voice interfaces.
English
@Prathkum AI made code sloppier.
Not the prompts — the missing verification step.
English
"Software isn't dead."
The new layer is software that controls AI.
That's where the next few years happen.
tae kim@firstadopter
Databricks CEO Ali Ghodsi: "If all software is dead, isn't OpenAI, Anthropic dead? They're software companies with a bunch of researchers writing software .. Nvidia should be dead because they just have really smart people who create chip designs, humans that use some software .. Software obviously isn't dead. Nvidia, OpenAI, and Anthropic are not going to be dead companies."
English
@firstadopter @CNBC Felt this. Ran one of the early BBS systems in '93.
Now building on Claude Code. Same vibe, 30 years apart.
English
Paul Tudor Jones on @cnbc “bought more AI stocks” “semiconductors” “It’s a crazy crazy time” brings up introduction of PC, Claude Code -> Microsoft 1981, Windows 95/internet. “beginning of productivity miracles that lasted 4-5 years” “we have a year or two to run” or “we continue to feel like 99” “October/November 1999” in terms of multiples. (either two years to run, another ramp to go)
English
@jrfarr And control before distribution is hard too.
English
