Vladislav Popov

AI agents in production, not in slides. PhD (York), EdTech founder, AI consultant. Open source: s57-parser, windows-admin-mcp. UK.

[Vladislav Popov @devladpopov]

@saxena_puru Cheaper tokens never meant smaller bills. It meant the agent now reads the whole repo to rename one variable.

[Vladislav Popov @devladpopov]

@cljack They stopped buying software and hired engineers to build it. Congratulations, they invented a software vendor with worse support.

[Vladislav Popov @devladpopov]

@DanKornas An agent that preloads every tool is like a plumber who brings the whole shop to change one washer. Then bills you for the van.

[Vladislav Popov @devladpopov]

@brexton Every year we find the best interface for coding agents. Every year it turns out to be whatever the last demo used.

[Vladislav Popov @devladpopov]

@Voxyz_ai /night and /morning is the first workflow where the developer finally gets eight hours of sleep and the reviewer gets none.

[Vladislav Popov @devladpopov]

The most useful line in my agents' logs is not the error. It's the name of the agent and the task it was on. Without that, every incident starts with an hour of "who did this".

[Vladislav Popov @devladpopov]

@monorkin Every Claude user eventually builds the same tool: a way to stop explaining yourself to Claude. The next one explains the first one.

[Vladislav Popov @devladpopov]

@cyrilXBT Opus plans, Sonnet codes, Haiku tidies up. The human pays the bill and reads the diff. Every org chart ends the same way.

[Vladislav Popov @devladpopov]

@ericwilliamrea From 5 PRs a week to 100. Somewhere in that building a reviewer is still doing 5, and he looks tired.

[Vladislav Popov @devladpopov]

@_Evan_Boyle 800 commits in 5 days is the impressive number. The honest one is how many of them a human actually read.

[Vladislav Popov @devladpopov]

@mattpocockuk AI didn't cure the tactical tornado. It gave him a fleet, a token budget and a dashboard that calls it velocity.

[Vladislav Popov @devladpopov]

What I'd copy even without Google:

1. One identity per agent, no shared keys
2. Every action logged with agent + task
3. Approvals tied to the agent

The model improves on its own. Accountability you build.

[Vladislav Popov @devladpopov]

Google's new agentic Gemini gives every agent its own Workspace account: own email, own context, actions logged under the agent's name.

Sounds like admin detail. It's the difference between an agent you can audit and one you can't.

[Vladislav Popov @devladpopov]

I run ~40 agents, one per project. The first real problems weren't model quality. They were:

Which agent changed this?
Who allowed that message to go out?
Why did two agents answer the same request?

All identity questions.

[Vladislav Popov @devladpopov]

@ChatGPT Every product now ships a feature you create from your phone. The hard part was never creating it. It was explaining to it what you actually wanted.

[Vladislav Popov @devladpopov]

@Jose_A_Alonso Math 2.0: the proof is checked by a machine, and the argument about whether it counts is still done by humans over coffee.

[Vladislav Popov @devladpopov]

@jomatech The first AI benchmark where the human baseline is also "leaves it to soak and walks away".

[Vladislav Popov @devladpopov]

@wheresryan22 Describe a machine in one line and get a rotating wireframe. Engineers spent decades on the opposite skill: describing it in three hundred pages so nobody can build it wrong.

[Vladislav Popov @devladpopov]

@Michael_J_Black A PhD advisor juggles six students who all believe their project is the main one. Agents are easier: at least they admit they forgot the context.

[Vladislav Popov @devladpopov]

Most agent failures I see are not model failures. They are permission failures.

The agent was allowed to act where it should only have been allowed to read.