Michael

Bring me the idea. I'll build it with AI. Software, automations and AI products for businesses. 27 years in business.

[Michael @arvanucom]

Question for anyone running coding agents overnight. In the morning, what do you open first, the diff or the review? And has that changed since you started?

[Michael @arvanucom]

@DivyanshT91162 Competing agents is fun but who decides which solution won? In my setup the agent that builds never signs off its own work, a different model does. If they all share discoveries, I'd worry they end up agreeing with each other instead of checking.

[Michael @arvanucom]

@MiguelPelaezAI Shared memory sounds great until one agent writes down a wrong assumption and the next one builds on it. I keep the builder and the reviewer apart on purpose so the reviewer doesn't inherit the builder's mistakes. Does it let you scope memory per agent?

[Michael @arvanucom]

@sumitmishra2026 Replays are useful. The thing I'd want highlighted is the retries. When an agent tries the same fix three times that's where the bad code usually comes from. Can it flag loops like that on its own?

[Michael @arvanucom]

@AMaajith0703 Seeing some of that on long sessions too. I ended up running agents detached in short jobs and reading the logs after, instead of watching the terminal. Less lag and you also catch when a run quietly stalls. Does it go away if you restart the session?

[Michael @arvanucom]

@wesleyboland Happened to me too. What's still worth building is the stuff around the agent, like who reviews, what it's allowed to touch, when it stops. The tools keep shipping features but nobody ships your rules for you.

[Michael @arvanucom]

Half my feed is glowing orbs and fire scenes from Muse Wonder and Opus 5.5.

Honestly I don't care how a 3D demo looks. What I want to know is whether a model gets through a long, dull job without me babysitting it. That's most of what businesses actually need built.

[Michael @arvanucom]

@rudraxx13 The 30% is the number I'd look at. Most small businesses aren't choosing between agents, they're not using any. In my experience it isn't the price, it's that nobody on staff can tell if the code is good. That's the gap worth closing.

[Michael @arvanucom]

@adrianhsuu Context loss is real, but I stopped picking one. Codex writes, Claude reviews, and the reviewer catches a lot of what got dropped. Splitting the jobs did more for me than switching tools ever did.

[Michael @arvanucom]

@eugeniu_ghelbur The nightly cleanup agent is the clever bit. Shared memory goes stale fast and then every agent trusts the wrong fact. How do you stop the cleanup agent from confidently rewriting something that was actually right?

[Michael @arvanucom]

@SciTechera Using two different agents is the part I'd copy. I have Codex build and Claude review, and they disagree often enough to be worth it. For something like a planet candidate I'd want to know which one caught what, and how many false ones got thrown out first.

[Michael @arvanucom]

One rule I don't break. The agent that builds something never signs it off.

Codex writes the code and Claude reviews it, on every push. I look after that.

Sounds slow but it isn't. The review runs while I'm doing something else.

[Michael @arvanucom]

@AravSrinivas Porting is where agents are honestly underrated. The old app is a full spec, so the agent always has something to check against. Greenfield work is harder because the spec lives in someone's head. Curious how long it ran before it was usable.

[Michael @arvanucom]

@AravSrinivas Low error rate on the sandbox is great but the errors that hurt me are the ones that succeed. The code runs, exits clean and does the wrong thing. Do you track anything beyond did it execute, like whether the result was actually used?

[Michael @arvanucom]

For a while my code reviews kept failing and I assumed Claude was the problem. It wasn't. I had set the review's turn cap too low, so it ran out of turns before it could finish.

Raised it to 60. I should have looked there first.

[Michael @arvanucom]

@rauchg Sandboxing gets the attention but the leaks I worry about are the boring ones. An agent with a real API key and a vague task does things you never pictured. I give mine the smallest key that works and log every outside call. What did the winning entries do about credentials?

[Michael @arvanucom]

Google says Gemini 4 Argon ran on its own overnight and made their ADK session prep 40% faster. I believe it. Codex builds for me while I sleep most nights.

The first hour of my morning goes to checking what it did. That part hasn't changed and I don't think it should.

[Michael @arvanucom]

@emollick Agree, and the lock-in is rarely the model. It's the prompts, evals and review steps people build around one model. I keep those so I can swap the model underneath in an afternoon. Most companies I talk to can't, because nobody wrote down what good output looks like.

[Michael @arvanucom]

@emollick Same conclusion here. My agents get their own inbox and anything going out as me stops for my OK first. The part I didn't expect is how much it helps the people on the other end. They answer the agent differently once they know. Do you let Dot reply without you?

[Michael @arvanucom]

The EU AI Omnibus was adopted this week and I'm already seeing "the AI Act is dead" posts.

It isn't. Annex III high-risk rules moved to Dec 2027. Transparency and AI literacy (Article 4) still apply.

If your staff use ChatGPT at work, Article 4 is about you. Now, not 2027.