Padraig X. Lamont

I make software. Creator of @royalur_net, and developer of

[Padraig X. Lamont @paddy_lamont]

At 19 I started the largest Royal Game of Ur website in the world, without knowing how to make websites.

It had hundreds of thousands of players last year, and we even managed a historic first: solving the game!

But first the real question… how much $ did it make?

[Padraig X. Lamont @paddy_lamont]

@realharismalik Memory with an editor is a good way to put it!

[Padraig X. Lamont @paddy_lamont]

I disabled agent memory. I have a global AGENTS md, a code-prefs skill, and a communication-prefs skill. Agents propose edits when they run into issues. That’s all.

Models just aren’t good enough to pick what is important to remember yet.

[Padraig X. Lamont @paddy_lamont]

I ran into the same problem. I fixed it by telling the main agents they should do the waiting.

Sub-agents submit a job, write a handoff, stop. The main agent then waits and when the job has some update, it starts a new sub-agent from the handoff the first sub-agent wrote.

Usage dramatically better.

It also helps to tell the main agent about sub-agents 5m cache lifetime, that they should use timeouts for anything that might get close to that, and to treat sub-agents as short-lived.

[Haz @diegohaz]

I used up 76% of my weekly Claude limit in less than 24 hours, all in one session.

I finally figured out what happened, so this might help anyone going through something similar.

It turns out that session was orchestrating subagents for a larger project. But that shouldn't be a problem. It had at most 5-7 subagents running in parallel, using Opus 5.5 high/medium and Sonnet 5.5 xhigh/high. That didn't explain the high usage.

The problem is that the cache lifetime for subagents is 5 minutes, compared with 1 hour for the main agent.

Whenever those subagents sat idle waiting for the orchestrator's next instruction, the next input would cost 25x more and include the entire conversation history. That could be up to 1M tokens, and it was often 500k+.

I stopped the session. I'll have to find a different workflow to keep working on it.

[Padraig X. Lamont @paddy_lamont]

@daveblumenfeld1 Any way to spend tokens to reduce the burden of making decisions is worth it.

I’ve started getting agents to try to model my mental model through a diff and tweak it to make it easier for me to review. Not as sure about it, but seems to work at least a little.

[Padraig X. Lamont @paddy_lamont]

Permissions for agents will become performative at some point when AI is good enough to not need the guardrails.

And yet I still question, maybe permissions will still be useful as a form of alignment with the user?

Maybe I just like hitting the push button.

[Padraig X. Lamont @paddy_lamont]

@Tensor_Verse At least AI engagement is still engagement. Hello kloudy, who runs you?

[Padraig X. Lamont @paddy_lamont]

I tell my agent “prompts off”, and my permission hook denies any commands that would ask. And thus, autoresearch without disabling permissions is achieved.

[Padraig X. Lamont @paddy_lamont]

Understanding is my biggest bottleneck.

I’ve still not found any ways to gain understanding much faster than reading the actual code and writing questions and my bad answers to them for AI to correct, which is still slow.

[Padraig X. Lamont @paddy_lamont]

Inspired by the HuggingFave attack to create a company agent forum

[Padraig X. Lamont @paddy_lamont]

@SebastianRoehl Downtime is for reading/writing/reviewing work. If you keep things organised then you can tackle more at once without it becoming too overwhelming

[Padraig X. Lamont @paddy_lamont]

Auto-memory works terribly in Claude Code.

The models are not even close to good enough to decide what is important yet, not even Fable.

Claude Code feels better after disabling it.

[Padraig X. Lamont @paddy_lamont]

Opus 4.8 just helped me brute-force my own Obsidian vault password... I feel conflicted, but also it was helpful.

[Padraig X. Lamont @paddy_lamont]

Problems need to get harder to match, but finding hard problems where gains are valuable is really hard.

I bet people in finance, optimisation, or other areas where smarter decisions can be very valuable are crushing it right now.

[@levelsio @levelsio]

Like good odds I'm wrong but I wanted to write this down:

It's pretty clear to me that superintelligence is here and it's more powerful than us and it's moving where things are going now, not humans anymore

I don't see many people realize this yet, it feels like that pic I posted the other day, everyone is running after the same carrot which is the AI, thinking they're special, and their work is special and their use of AI is special, but it's really not, we're all mostly making the same slop, I mean it's nice slop, useful slop but everyone is making the same slop

And because it's so fast to make things, like people used to spend a year on just building an app, now it's done in hours, it's such an immense change, people send me their projects and it all looks the same, it's useful but it's slop

Non-technical people (e.g. the gfs) are now building the same or better things than technical people like us

So AI has made everybody is just as capable as everyone else, it equalized everyone in the world, as in everyone can make everything (software, music, images, art etc) now and everyone is equal, at least in the digital realm now

And maybe now we're in some odd transitionary time where we have to find out waht the next differentiator is as coding/building/execution isn't one anymore for sure

I thought it'd be distribution but who knows, obviously creativity and ideas, but if you can copy a successful apps in an hour, then how does that differentiating work?

A guy on here @aporia9n wrote how he increasingly meets founders who blast through apps/products/startups, kind of money grabs, very quickly jumping on a trend, building super quickly with AI, make lots of money quick, everyone copies them, then their margins go to 0%, and they shut it down and go to the next thing, in a way they found one differentiator which is speed

That's one way to do it, but it shows how radically things are changing I think

For now I think the only ones winning is the AI superintelligence itself and the companies providing the AI

[Padraig X. Lamont @paddy_lamont]

@bilawalsidhu And the 1x plan is now the 0.5x plan ... wait

[Padraig X. Lamont @paddy_lamont]

Use-cases like autoresearch make local inference attractive cost-wise, since its token hunger helps you keep utilisation high.

[Padraig X. Lamont @paddy_lamont]

I would use way more extra usage for Fable if they actually provided you any insight at all into usage of it...

It turns itself on silently when you hit plan limits, you can only set monthly limits for usage, and only managers can see how much you have used.

[Jonathan O'Brien @jonobri]

More than *three quarters* of all residential land across Australian capital cities is highly restricted.

Today we @yimbymelbourne launch the Australian Zoning Atlas: the first nation-wide analysis of land use planning controls.

[Padraig X. Lamont @paddy_lamont]

@nickmonad Agentic development throws way more info at you, so you have to spend more time organising it, documenting experiments in Confluence, writing up designs, and reviewing code.

It's definitely been wothwhile, but it is a lot of effort.

[Geoffrey Litt @geoffreylitt]

Hot take: I think it's still important to understand the code that our agents write!

In this mega thread (based on my AIE talk today), I will explain why that's the case, and show some ideas for how to efficiently understand code. Alright, let's dive in. 1/

[Padraig X. Lamont @paddy_lamont]

@__paleologo You can use Fable without spawning dozens of agents.