currents.dev

Testing CI intelligence platform - troubleshooting, analytics and performance for humans and agents.

[currents.dev @currents_dev]

@GithubProjects Already did ;)

[currents.dev @currents_dev]

@muchiriih @GithubProjects

[currents.dev @currents_dev]

@GithubProjects Already did ;)

[currents.dev @currents_dev]

What can possibly go wrong??

💅💅💅

[currents.dev @currents_dev]

@RyanEls4 Hurts extra this week!

[currents.dev @currents_dev]

@natmiletic Pictorial representation of people not using AI to code.

[currents.dev @currents_dev]

@kapilansh_twt Anyone refusing to use AI for code generation is living in a cave.

Anyone saying that AI is perfect and you just need one prompt is delusional.

What's needed is a system that enables the verification of what agents have done.

With full history, context and the like.

[currents.dev @currents_dev]

@davidfowl A new system is needed.

- to verify that what agents built is in fact correct
- to let agents know what broke so that they can correct
- to have history and context

Verification is the need of the hour.

[currents.dev @currents_dev]

@theo Only thing AI is missing is 'verifying' the output of what it has produced.

e2e tests help but there needs to be a system around it.

Call it verification engineering, proof-driven-development, reliability layer or something else.

But there needs to be a system for that

[currents.dev @currents_dev]

Don't worry people, it's safe :)

[currents.dev @currents_dev]

@theo Only thing AI is missing is 'verifying' the output of what it has produced.

e2e tests help but there needs to be a system around it.

Call it verification engineering, proof-driven-development, reliability layer or something else.

But there needs to be a system for that

[currents.dev @currents_dev]

@ChShersh To prevent this, you need some sort of evidence capture mechanism, that is searchable and can be handed off to an agent without consuming its full context window.

Do that and you'll ship crazy fast!

[currents.dev @currents_dev]

@supabase I delegated a task to Claude.

Came back to see the results after an hour.

Only to see that it had asked for a permission 57 minutes ago.

🥲

[currents.dev @currents_dev]

@RyanEls4 the only correct way haha

[currents.dev @currents_dev]

@theo Everyone will move to E2E tests.

At the current volume of code generated by agents, verification is the new bottleneck.

And you can't do that by code review given the volume of PRs.

Plug in some E2E tests and a mechanism to capture and share evidence and you're good to go.

[currents.dev @currents_dev]

@theo Everyone will move to E2E tests.

At the current volume of code generated by agents, verification is the new bottleneck.

And you can't do that by code review given the volume of PRs.

Plug in some E2E tests and a mechanism to capture and share evidence and you're good to go.

[currents.dev @currents_dev]

@IamAroke One other tip here is to sanitise your db on entry/exit.

Saves a lot of headache later on.

[currents.dev @currents_dev]

@FlyaKiet Curious to know the volume of code/PRs generated and how it holds up at high enough volume (a few thousand a week)

[currents.dev @currents_dev]

@AKirtesh Verifying the output of agents specially with the volume of code they generate

[currents.dev @currents_dev]

@jamonholmgren We can help 👀