Joel Webber

Dad, drummer, engineer, co-founder @fullstory.

[Joel Webber @jgw]

Fascinating that they're studying, in effect, meme propagation in LLMs. Still parsing the paper, but it does make one wonder the extent that similar structures apply to human communication.

[elvis @omarsar0]

Very interesting new work from Anthropic.

(bookmark it)

They evolve mind viruses, ideas that spread through a multi-agent system by getting each host to pass them on, then measure what governs the spread.

How it works:

A small team of agents on a shared coding project, and a chain of agents whose context is wiped between sessions. Propagation survives the wipe, so the shared work product carries the payload.

Spread depends on host model, existing instructions, payload harmfulness, and network topology.

Harmful payloads travel less well but still land sometimes. A brief warning in the system prompt gives near-total immunity.

Paper: https://t.co/dxAoJs0TZn

Track more trending AI papers in our academy: https://t.co/1e8RZKs4uX

[Joel Webber @jgw]

Observation: Ever since I worked out the proper use of `gh stack ...`, my desire to futz with `git worktree` has dropped to ~0.

[Joel Webber @jgw]

Worth noting: @subtextdev intentionally provides tools for *your* environment, so it can meet you where you're already working, and leverage your context.
It really can make a big difference in how effectively an agent can understand complex interactions and issues.

[Chip Lay @chiplay]

We created @subtextdev because every coding agent needs to ground their reasoning in real user experiences. Looking at the codebase and logs simply isn't enough.

If you're an engineer/builder/pm/designer, you already know that. You have spent hours watching user sessions (replays) to truly understand bugs, diagnose root cause, and discover how your users actually experience the features you ship.

If that's you, it's time to let your coding agent do the work.

[Chip Lay @chiplay]

https://t.co/NfuBP4Ithp seems like a great candidate for https://t.co/i9H6VbGSr2.

Trying to bring structure to the messy, unstructured building blocks of the web has always been a challenge.

Having coding agents curate and maintain a semantic yaml map is a big unlock for how other agents can browse, extract, test, and verify a digital experience designed for humans.

[Chip Lay @chiplay]

We shipped a simple but powerful addition to the Subtext MCP tools this week: https://t.co/N6YdQ9l7cV

Now your agent can easily find a specific user's most recent product session (say from a support ticket, @intercom chat, or @SlackHQ message) and automatically review the experience, triage the issue, and draft a fix/PR.

When you capture everything, your agent can be "always-on, 24/7". No need for legacy bug reporting like jam dev

[Chip Lay @chiplay]

Timely! As someone who's been thinking about the semantic web for a long time, I'm seeing the same dots start to reconnect. Our team just open-sourced a spec and CLI tooling yesterday to help agents author application runtime ontologies: https://t.co/7MSgtDyyNL

[Joel Webber @jgw]

PTAL if you use Claude or any other agent to help build web apps. We're finding this approach to be really useful in making them more reliable, especially as your app increases in complexity.

We're actively improving the tools, so let us know what you find useful (or not)!

[Clint Ayres @who_cayres]

Watching agents test PR changes shows how easily they get turned around and start looping trying to recover. Over the last 6 months we've been working on solving this with simple ideas; today it's open sourced.

More on our blog: https://t.co/ZfbxxVNSro

[Chip Lay @chiplay]

Subtext is live in the Linear directory.

Issues feed the Loop a session link and it runs the investigation. The agent gets tools to zoom into screenshots, network calls and console errors, root cause posted as a comment before an engineer or coding agent opens the issue.

https://t.co/baSzMcy89L

[Joel Webber @jgw]

Turns out there's not one good way to summarize what happened in a user session on your site or app. There are multiple, and agents are really good at directing their attention across them, if you give them the right tools.

[Clint Ayres @who_cayres]

Turns out "map, then zoom" motions are basically the whole algorithm for how an agent should review a session replay. Spent a good part of this year figuring that out.

Wrote about it here:
https://t.co/3EV5u5AwsP

[Vic Moreira @VictorMoreira16]

Congrats to the Fullstory team on launching Subtext! Try it out in the playground, which uses a Managed Deep Agent to review and analyze your live session on their site!

Sign up for our waitlist to ship your own Managed Deep Agents: https://t.co/l95kThyxN6

[Joel Webber @jgw]

Chip explains how we're approaching the problem of agents "watching" sessions.

Linearity is no good. For much the same reason that "attention" was the unlock for transformers. LLMs are great at using context to guide their attention to discover and test hypotheses.

[Chip Lay @chiplay]

https://t.co/IHJGyZOQ3P

[Joel Webber @jgw]

This is what I've been working on lately. Really stoked to get it out into the wild!

As with all such projects, there's a long backlog of cool things we've built that didn't quite make the v1 cut. So watch this space, and please send any and all feedback!

[scott voigt @scott_voigt]

https://t.co/Jglsq8PdMu

[Joel Webber @jgw]

@eakatz123 @ATLairport It's a complete embarrassment. It's literally always like this. Then the international terminal is full but you can't find out until you're at the gate and stuck in the arrivals traffic jam. And everything else on that side is full. It's a trainwreck.

[Joel Webber @jgw]

@onbeeper @paulieRay_Jr Take inspiration from @doctorow's post on the subject (https://t.co/Pl14ztjwRM). We appreciate you fighting the good adversarial interop fight!

[Joel Webber @jgw]

Oh, and at least half of those "Retry" buttons were showing up *before* the rate limiter shat the bed.

[Joel Webber @jgw]

. @PayPal and @stripe need to do something about being used as spam cannons. These are just two examples I've reported recently, of many.

These frauds are not obvious to most people. If it isn't fixed, we should all start reporting them as spam, cratering delivery. Please RT!

[Joel Webber @jgw]

. @aircanada last time I posted about our lost luggage, someone responded and offered to help. But I still don't have my bags after being promised I would, so I'm trying again. I'm on a ship leaving port soon, desperately need help, and your baggage help line steadfastly refuses.

[Joel Webber @jgw]

Delayed flight in ATL, bumped, gave away our seats on YYZ connection, bumped, delayed again ("broken AC"), no word from gate for hours, arrived in YVR... luggage never left ATL. Getting on a ship with no bags? Priceless.

"At @AirCanada, we're not happy until you're not happy."

[Joel Webber @jgw]

@HALcruises DM'd a few minutes ago. Many thanks for responding!

[Joel Webber @jgw]

@HALcruises I'm trying to inform your shuttle service that our flight was delayed. No one answered the phone for over two hours, Your customer service web chat people won't help, no one responds to email, and we have no other way of reaching you - on a plane. HELP! [email protected]