Ville

Staff engineer focused on AI augmented engineering, AI literacy and Developer Experience @datadoghq · Father to two small humans

[Ville @efexen]

What % of your swe work do you think sota models can do well today?

And what % of your colleagues work?

Just curious 🤷‍♂️

[Ville @efexen]

I spent the last 6 months or so pretty Claude pilled both work & home despite them messing with Betty (my OpenClaw) back in April

Spent last couple of weeks trying codex and gpt 5.5 and it’s just simply better, by a lot

If you’ve not tried switching for a while I’d urge you to

[Ville @efexen]

https://t.co/6IXZfRjzxB

[Ville @efexen]

An intent record isn't just for humans deciding what to build.

It's institutional memory in a form an agent can actually use later: the original signal, supporting evidence, the decision, and the reasoning.

Agents need to know not just what exists, but why it exists.

Tickets don't convey that. Intents could.

[Ville @efexen]

A lot of teams don’t have a dedicated PM, PO, or delivery manager.

The work still happens. Engineers absorb it: tracking signals, inferring priorities, and holding loose threads in their heads.

That work is invisible until it breaks.

An intent layer doesn’t add ceremony to it. It just makes it explicit.

[Ville @efexen]

Your session “From Vibes to Data: Evaluating Agents on Your Real Work” is successfully submitted!” 😬

First big conference CFP for @AgenticAIFdn AGNTCon + MCPCon EU sent

Why are my hands sweaty already?!

[Ville @efexen]

The new chatgpt personalisation thing recommended something for me to do this weekend and then concluded:

That’s peak “high-agency dad-engineer” energy.

👌

[Ville @efexen]

One overloaded review gate at the end is the worst place for product judgment.

Better: a few lighter gates at better moments.

Should we build this. How should we build it. Did we build it correctly. Did it behave as expected.

Same total decisions, much earlier, much cheaper to change.

[Jason Fried @jasonfried]

Bragging about how much software you’re shipping with AI is like holding down the shutter button and bragging about how many photos you took.

[Ville @efexen]

The “AI-native SDLC” conversation is too focused on coding agents.

The bigger unlock is upstream: ingest a messy signal, cluster related evidence, and prepare a concise case for human review.

The interesting agent isn’t the one writing code. It’s the one preparing the brief that decides whether code should be written at all.

[Ville @efexen]

An intent isn’t a ticket. It’s evidence that something might need to change.

Cheap to capture. Numerous. Most never become work.

The point is they’re preserved in a form you can reason over later, not lost in Slack.

Tickets track work. Intents track the case for work.

[Ville @efexen]

@steipete This would be a lot more impressive if the quality of the releases wasn’t plummeting

[Ville @efexen]

OpenClaw has broken core behaviour in nearly every recent update

Whether that’s because humans stopped looking or because the system outgrew the people looking at it 🤷‍♂️

Once approval stops tracking comprehension velocity is just the speed at which quality drops

[Ville @efexen]

A ticket is a commitment.

Someone named it, put it on a board, and made it part of the team’s inventory.

That’s too heavy for “this error message confused me.”

So we either skip capturing it or create a low quality ticket nobody trusts.

Both are bad!

The fix isn’t better tickets it’s a layer before tickets 🤔

[Ville @efexen]

@thekitze 😰

[Ville @efexen]

Most signals that something needs to change quietly disappear.

Slack threads, customer comments, screenshots in DMs, or engineers noticing repeated confusion.

Some become Jira tickets too early. Some live in someone’s head.

Some get rediscovered three months later as if they were new.

[tobi lutke @tobi]

https://t.co/5MJ7u1VHwf

[Ville @efexen]

Code review is the only review most teams do.

Often it’s where we ask: is this a good idea, is the design right, is the code good, are tests enough, is it safe to ship.

That’s a lot!

No wonder reviews are slow and emotional.

By then you’re already negotiating from sunk cost.

[lossybrain @lossybrain]

The limiting factor for AI software engineering is just as likely to have been a problem before the advent for AI, it's just now bigger because of AI. https://t.co/J5cKHv4y9a

[Ville @efexen]

How active is the Agent Trajectory Interchange Format (ATIF) from Harbor?

Looking for a common format to use across different harnesses for evals platform.

Anyone using it or have other recommendations to look at?