Most AI meeting assistants are solving the least interesting part of the problem.

They transcribe the call, write a summary, spit out action items, and leave you to do the actual work.

I don't need another to-do list.

If an AI understands what's being discussed, why can't it rk?

That's what we're building with Skyflo.

In this demo, Marin joins my meeting, speaks, shares its screen, and coordinates agents working across five repositories.

I'm in a rush, so Marin suggests moving the mission to Skyflo Cloud. It identifies a release blocker, resolves it on its own, and keeps the agents working after I've left the meeting.

And the best part? Those agents are building Skyflo itself.

Skyflo builds Skyflo.

The future of AI isn't another assistant that waits for instructions, answers questions, and hands everything back to a human.

I want agents that understand the context, coordinate with each other, make progress independently, and actually take responsibility for getting work done.

Taking notes was never the hard part. Everything that happens after the meeting is.

Skyflo manages its own outbound, end to end.

Research every hour. LinkedIn every 30 min. Inbox every 10. Meetings answered the same day. CRM kept current. A brief for me at 6 every morning.

Skyflo sells Skyflo. Skyflo markets Skyflo. I take the calls.

The new CoD campaign is the best till date.

Even better than the OG Modern Warfare.

There, I said it!

Opus 5.5 is still the best model out there.

I don’t care if some leaderboard got improved with a newer model. Most of these launches don’t change what it’s actually like to hand work to an agent and trust the output.

I’ve been using the newer models. Some of them look good in demos and on certain benchmarks. Then you put them in a real codebase and they confidently invent how things work. That’s not a minor quality difference. It’s the difference between something that reduces work and something that creates more of it.

Opus still feels like the one I can actually delegate to without constant babysitting. The others keep proving they’re better at sounding capable than being reliable.

Until that changes, the model rankings are mostly noise.

Meta's Muse Spark Open is an interesting moment for companies that want to run their own AI.

The model has 816B parameters, with 42B active per token. Apple says its 4-bit weights take about 493GB before runtime memory, and has published a setup that splits it across two Mac Studios.

That's an office infrastructure decision. You can put a model on dedicated machines, keep inference on your own network and decide when to upgrade it. You also take responsibility for utilization, power and keeping it running.

One timing detail: Apple's announced date for the 512GB Mac Studio configuration is late October. The demo is ahead of that hardware's announced availability.

The useful question for a company is which work deserves its own hardware. A constant workload involving sensitive internal data makes a very different case from an agent used occasionally throughout the day.

Skyflo’s reviewer can inspect a change and report findings. It cannot edit files.

A reviewer that quietly fixes the work becomes another implementer. Keeping those jobs separate makes the findings easier to trust.

It's honestly pathetic what Google shipped with Gemini 4 Argon.

They have the chips, the talent, the data, the distribution, and basically unlimited capital. They should be one of the hardest companies in AI to compete with.

Instead, months of delays, endless benchmark flexing, a 1M token output limit, and then they lock their flagship model behind Ultra in their own agentic coding tool.

And the paywall isn't even the biggest problem.

I couldn't care less how many benchmarks a model tops if I can't trust it to do actual work. An agent that confidently makes things up about a codebase isn't saving me time. It's creating more work. At that point, what's the point?

Google seems more interested in proving how intelligent its models are than making them genuinely useful.

The irony is that they have almost every structural advantage imaginable, yet they keep making decisions that give smaller companies room to compete.

You can't benchmark your way into developer trust.

Deleting a button can make an app cleaner and harder to use.

I want fewer repeated controls, not fewer clues about what to do next.

I’m building Skyflo because I think AI should be able to take on far more of the work of running a company.

Not just answer questions or complete individual tasks, but actually carry work forward.

Think about how much of a founder’s day goes into connecting things. A customer oduct priorities. Research informs a decision. A launch involves engineering, marketing, distribution and follow-ups.

Context gets scattered across tools, and someone has to keep track of it all.

That's the problem I'm solving with Skyflo.

Today, Skyflo builds Skyflo.

I'm not just talking about writing code. Through Skyflo, agents build and maintain our desktop and mobile apps, frontend dashboards and backend systems, and manage everything from deployments and release rollouts to cloud operations.

I set the direction, make the important decisions and review the outcomes. The agents coordinate and execute the work.

But engineering is only one part of what I'm building.

Skyflo already gives agents persistent context, shared objectives, access to tools and the ability to work together across environments, with humans retaining control over the decisions that matter.

As Claude, ChatGPT, Cursor and other tools improve and become capable of handling more work across engineering, design, sales, marketing and beyond, Skyflo becomes more capable alongside them.

You shouldn't have to rethink how your company operates every time a better model comes out or a new tool becomes available.

Ultimately, I want to reach a point where running a company doesn't mean personally coordinating every moving part.

You set the direction. Skyflo carries the work forward.

That's what I mean by "the agentic operating system for your company."

A payment handler can say “already processed” while the customer still has no access.

In this crash-and-retry repro, the receipt commits before the seat grant. Every retry then skips the grant.

In one DB, commit both together. Put that invariant in the agent brief.

Has anyone else noticed how bad the new ChatGPT desktop app has gotten?

The UX has regressed, and the performance is ridiculous. Mine hit 18 GB of memory before macOS threw an out-of-application-memory warning.

How does a company like @OpenAI ship a desktop client this bad?

A Skyflo test on a real iPhone got all the signature checks right and still asked someone to approve a command they could not see. The screen said “Bash.”

The fix brought the command and file changes into the approval itself.

For remote agent approvals, test what the person can e phone screen. A valid signature cannot fill in missing context.

A small Python cancellation repro: both tasks run their cleanup. The one that re-raises CancelledError ends cancelled. The one that catches it and returns ends with a normal result.

That difference reaches the caller. Put cleanup in finally; when you catch cancellation to clean it afterward.

Replacing a file changes what the path points to. It does not refresh readers that already opened it.

In a small local Python repro, a child writes a temporary file, replaces the original, then exits. The reader holding the old file sees OLD; a reader reopening the path sees istinction for config reloads: atomic replacement and fresh reads solve different problems. This tests namespace replacement, not power-loss durability.

A small SQLite WAL repro: one connection reads 0, another commits 1, and the first still reads 0. Trying to write from that old snapshot returns SQLITE_BUSY_SNAPSHOT.

Ending the read transaction makes the new value visible. A retry needs a fresh read before it recomputes the write.

Today a reply published, but our verification check couldn't match its link. X had wrapped the URL across lines.

We checked the exact permalink and the link's destination. Same copy, same parent, same account. No second click.

After a send, 'I can't verify it' and 'it didn't recovery paths. Otherwise a formatting mismatch becomes a duplicate post.

The image model drew our 1.4 startup bars at 35% and 59% of scale. The numbers say 32% and 51%.

On a performance card that's a false claim, however good it looks. We fixed it to the pixel before it went anywhere, and every data card now gets its bars measured first.

A harness-parity bug we fixed in today's Skyflo release.

With review on, Codex asked the person before every call to Skyflo's own tool gateway: reporting, file search, listing missions. Claude Code made the same calls without that prompt, because the gateway already checks the call's capability token and the tool's policy, and asks you itself when an effect needs it.

Two gates on one agent, one on the other. The Codex adapter now answers approvals for our own gateway only. Every other MCP server's approvals still come to you.

Replit's harness post has the number I'd want: the same setup with one long-lived worker scores 11 and 16 points lower than a core loop that dispatches briefed subagents.

That's how I'd build it too. One loop holds the plan, and each worker gets a brief instead of inheriting the whole thread.

Near-Astra at a fifth of Astra's price is only a fifth of the cost if 6.1 Sol doesn't spend more tokens getting there.

That was my problem with GPT-6 Sol. On long agent missions it overthought, turning work Opus finished in under an hour into multi-hour expeditions.

The number I want is tokens per finished task.