Alex Cinovoj

@zocomputer Let's go, team! I love this future!

JUST SHIPPED: working with files on Zo should feel smooth as butter now 🧈 We made a nested sidebar to help you navigate across files seamlessly. Also Zo can read any file now, from PDFs (commenting + highlighting) to Word docs, spreadsheets, ppts, audio, and more. Try it today!

OpenAI just said the quiet part out loud about agent spend. Token price is becoming the wrong scoreboard. Their new enterprise guidance points at “useful work per dollar”: tasks completed, time saved, decisions improved, and workflows ready to scale. That is the right frame. A cheaper model that retries, loops, or needs human cleanup can still be the expensive option. This is where most agent programs get sloppy. They buy capacity before they know the accepted outcome. They track credits before they track completion rate. They raise limits before they define who approved the action, what context was used, and where the rollback lives. If the agent cannot prove the work it did, the bill is not an investment signal. It is just smoke from the machine. Cost control starts with receipts. What would you measure before giving an agent more budget?

Agent rollback is where the demo becomes a system. Most teams can show an agent completing the happy path. That is not the hard part. The hard part is what happens when the tool call is wrong, the memory pulls stale context, the handoff misses the owner, or the agent keeps running after the business rule changed. A production agent needs a receipt chain. Who approved the permission? What did it read? Which tool did it call? What output was accepted? Who owns the rollback when the answer is technically valid but operationally wrong? If that path is not written down before launch, the launch is mostly theater with better screenshots. The model can be impressive and still leave the company exposed. The control plane is the product. What breaks first when your agent is allowed to act without you watching?

GitHub Issue Fields look boring. That is why they matter. Agents fail when work arrives as a pile of prose. Someone has to infer priority, effort, owner, component, target date, risk, and what “done” means. That is where handoffs rot. Typed issue metadata turns the ticket into a control surface. Now an agent can triage the work, route it, filter it, and leave a receipt the rest of the system can understand. The model is not the release gate. The fields around the work are. What field would your agent need before you trusted it with the next ticket?

@zocomputer @0thernet One of my favorite tools lately

Why did @0thernet build Zo? Why is it so important for everyone to be able to own their own land on the Internet? Why does he let his agents run 24/7 like a madman? Over the next few days we'll lore-drop more about how and why Zo Computer was built. Stay tuned 🎬

@Xentrilo Agreed.

@AlexCinovoj 5.6 is a step up for sure, the gap is much less of a jump compared to the Grok 4.5 jump, but still the toolkit now is crazy.

ChatGPT Work is going to expose a boring production problem. OpenAI can ship a better agent, three GPT-5.6 models, mobile task control, and reasoning knobs. The user still has to know which model to pick, how much reasoning to spend, when the agent should run, and what counts as a good result. That is where most AI rollouts quietly break. Not at the model layer. At the operating surface. If a non-technical operator needs a translation table before they can use the system, the agent is not finished. The missing product is the release gate around the work: defaults, receipts, fallback paths, and clear ownership when the run goes sideways. The next fight is not who has the smartest model. It is who makes the smart model boring enough to trust on a Tuesday afternoon. What control is missing?

GPT-5.6 quietly moved the agent fight into the control plane. The launch headline is bigger models, better coding scores, and more work per token. Fine. Useful. But the operator signal is lower in the page: programmatic tool calling in the Responses API, beta multi-agent requests, and explicit cache breakpoints. That is where agent systems start to look less like smart chat and more like production software. Can the model coordinate tools without leaking state everywhere? Can it spin up focused subagents and synthesize the work without making the owner babysit every step? Can teams control cost and cache behavior instead of praying the invoice behaves? Benchmarks get attention. Control surfaces decide whether the thing survives Monday morning.

Most agent failures are ownership failures. The model did what it was told. The workflow ran. The demo looked fine. Then nobody knew who owned the permission boundary, who could pause it, where the logs lived, or what rollback meant after a bad run. That is not an AI problem. That is an operating surface problem. Production agents need named owners, release gates, durable logs, access revocation, and a boring handoff path before anyone should trust them with customer-facing work. Real systems need boring controls before clever behavior. I do not care how sharp the prompt is if the system has no receipt chain. The difference between a demo and an agent is not autonomy. It is accountability. When it drifts, who gets paged?

Most agent failures start as permission problems. The team gives a model access to tools, files, inboxes, calendars, CRMs, and half the company knowledge base. Then they call it an agent because it can take action. That is not a system. That is a loaded shortcut with no receipt chain. The boring controls are what make it usable. Time-boxed access. Owner-named approvals. Tool scopes. Run logs. Rollback path. Kill switch. A handoff record when the agent gives the work back to a human. If those pieces are missing, the first clean demo just bought you a future incident. The model is maybe 10% of what makes an agent work. The permission layer is where the risk lives. What control would you refuse to ship without?

@miltonheyan 100% this is the way!

@AlexCinovoj the model was never the bottleneck — the control layer is. that's the whole bet. and it's not evenly boring: logs + rollback are the easy 80%. the hard 20% is a gate that stops the run before a move you can't walk back. we went all-in on that → x.com/CoreSpeedHQ/st…

Most agent projects don’t fail because the model is weak. They fail because nobody built the boring control layer around the agent. A tool-using agent needs logs you can trust, permissions you can explain, rollback when it goes sideways, and a human owner who knows what “done” means. Without that, you don’t have an AI teammate. You have a demo with production access.

The next UX shift is simple: agents do not need fewer buttons. They need more capable systems.

@domwhyte42 When agents become the user, “simple” stops meaning fewer features. It means clean structure, reliable actions, and enough depth for the agent to actually get work done.

Simple software is dying. Simplicity only mattered when humans had to use every feature. Now that agents can operate software for you, it’s the most capable tool that wins. Fillout reached 12,000 paying customers by focusing on depth over simplicity. It turns out we may have also built the best form builder for agents. Meet the new Fillout:

@BunsDev Good luck!

I’m stepping away from #OpenClaw Within 6 months, I shipped Control UI v2 and ClawHub v2 as a FOSS maintainer. I’m incredibly proud of my work and grateful to those who believed in me 🦞 The best part of any ending is what it makes room for. Time to build what’s next… 🌙

New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

Cloudflare just made crawlers pay. That is the web changing shape. Free extraction used to be the hidden subsidy behind search, training data, and a lot of agent products. Now the gate is visible. If your AI workflow depends on crawling, the cost model just got a new line item. The question is not whether the crawler still works. It is who controls the access layer. What data source are you still treating like it is free forever?

The agent is not the product. The control surface is. Everyone wants the agent to click buttons, write code, move files, call APIs, and make decisions. Fine. But the expensive part is what sits between the agent and the blast radius. Right files. Right token. Durable log. Human stop. Rollback path. That is the boring layer buyers actually need before trust. That is where pilots turn into production. Not in the prompt. Not in the benchmark. In the permission path. What control is missing?

Meta just gave everyone a production warning. Reportedly, leadership admitted AI agents did not move as fast as expected after a massive reorg. That is not a Meta story. It is an enterprise AI story. Most teams still treat agents like a talent problem. Hire sharper people. Buy stronger models. Move more teams into AI. Then Monday shows up. The agent needs permissions. The workflow has exceptions. The data is stale. The owner is unclear. The rollback path is missing. The logs only show the happy path. That is where the pilot dies. The model can be smart and the system can still be useless. Agents do not fail because the demo was bad. They fail because nobody designed the operating surface. Identity. Access. Evals. Observability. Kill switch. Human escalation. Owner-named controls. Boring list. Expensive to skip. If your agent can act without a receipt chain, you do not have automation. You have a risk surface with a nice chat box. What breaks first when your agent leaves the demo?