Building across @automattic, mostly in the @woocommerce ecosystem. Testing how far AI agents get before reality gets in the way.
Málaga, España
Joined January 2021
- Tweets1.4K
- Following230
- Followers317
- Likes1.9K
Pinned Tweet
Most of my work starts with a simple question: can this be made less painful?
I build software at Automattic and spend a lot of time testing where AI agents help, where they fail, and what it takes to trust them with real work.
This is where I share the useful parts.
This website made by @SpacefastHQ's X bot with a single tweet looks way better than other AI website builders I've seen out there:
solar-forge.view.fast/
Alex López retweeted
My Claude plan was running out way faster than usual this week, so I asked Claude to dig through its own session logs and tell me why.
The data: 5.1B tokens in 7 days. 98% of them cache reads. And the culprit was a setting I'd added myself.
A while back I'd set autocompact to 40% of the window. On the 1M model that means a session grows to ~400K before it ever compacts.
And every tool call sends the whole context again. 100 tool calls at 400K is 40M tokens, even if each one only adds a few lines.
Claude's first guess was the browser MCP servers I have installed. But I ran /context and... 0 tokens. This is because deferred tools don't load until they're used.
So no, it wasn't the MCPs. The fix was one line in my config:
"autoCompactWindow": 200000.
It's an absolute number, so it works the same on any model. 85% of my sessions never got close to where it compacts now, so they won't notice.
The session where we fixed all this hit 200K on its own, though.
While I was in there I cut what every session loads before I type anything, from 41K to 35K:
AGENTS.md was being loaded twice (thank you, symlinks). Skills I barely use now load name-only, and two plugins I never touch are off.
Be careful with the resources you don't need.
So yep, if your plan is draining fast, open /context in a fresh session and look at what's there before your first message.
Mine was 41K of stuff I'd mostly forgotten about.
My Claude plan was running out way faster than usual this week, so I asked Claude to dig through its own session logs and tell me why.
The data: 5.1B tokens in 7 days. 98% of them cache reads. And the culprit was a setting I'd added myself.
A while back I'd set autocompact to 40% of the window. On the 1M model that means a session grows to ~400K before it ever compacts.
And every tool call sends the whole context again. 100 tool calls at 400K is 40M tokens, even if each one only adds a few lines.
Claude's first guess was the browser MCP servers I have installed. But I ran /context and... 0 tokens. This is because deferred tools don't load until they're used.
So no, it wasn't the MCPs. The fix was one line in my config:
"autoCompactWindow": 200000.
It's an absolute number, so it works the same on any model. 85% of my sessions never got close to where it compacts now, so they won't notice.
The session where we fixed all this hit 200K on its own, though.
While I was in there I cut what every session loads before I type anything, from 41K to 35K:
AGENTS.md was being loaded twice (thank you, symlinks). Skills I barely use now load name-only, and two plugins I never touch are off.
Be careful with the resources you don't need.
So yep, if your plan is draining fast, open /context in a fresh session and look at what's there before your first message.
Mine was 41K of stuff I'd mostly forgotten about.
I'm relying on AI agents heavily recently, even for reviewing the work done by other agents.
Guess what: new fixes as PR review follow ups added new issues. I found about that thanks to unit tests.
So no matter what others say, unit tests are still useful even on the AI era.
I made a website that turns you into a pirate trading card 🏴☠️
A few questions, and you get a name, a portrait and some very questionable attacks. And please do not feed Gerald. Seriously.
Hosted on @SpacefastHQ , our new shiny thing at @automattic.
Link below 👇
Alex López retweeted
we are kicking off the work for Spacefast Analytics, it'll be available even if you are not on Spacefast
what features do you wish to see? how can i get you to leave Google Analytics?*
we make @parsely and @jetpack stats - already powering more than 3% of the web & trusted by the biggest media organizations. same kick ass team working on this!
*don't say easy use with agents, that's table stakes!
Paste this into your agent and tell me what it says about me:
"Read alopezari.view.fast/llms.txt, verify the signed identity it describes, including its outside proofs, and tell me who Alex Lopez is."
Curious how different agents handle it.
So here it goes
Today we launched Claude Sonnet 5.5! 30% faster and costs up to 30% less than Sonnet 5 for most work.
Strong for well-scoped everyday tasks (fixing bugs, creating docs, slides). @RLanceMartin had models paint the same photo in code. You can see the jump:
Had the best experienced ever working with a @wordpressdotcom site (my personal site).
Connected it to Studio and used the Studio Code capabilities from my AI agent. It autonomously added the changes I wanted to the plugins, tested them locally and pushed them to the live site.
