2026-07-05 · Why implementing ActivityPub is hard, and why it doesn't have to be

Transcript

Transcript. Paraphrased; sources in notes.md.

HostWelcome to July 5th, 2026. I am Daniel, and joining me today is Samantha. Samantha, I saw you were already digging into some of today's topics before we hit record. How are you holding up?

GuestDaniel, I am fully caffeinated and ready to rumble. I have been deep in the weeds with agentic workflows all week, and some of the stuff coming out of the Lobsters community today is hitting home in a way that is both refreshing and slightly terrifying. It is that perfect mix of building things and wanting to throw your laptop out the window because of a hidden schema error.

HostThat is exactly the energy we need. We are pulling all of today's topics from Lobsters, the tech community where engineers actually go to hash out the details. Let us start with ActivityPub, the protocol powering the fediverse. There is a new post arguing that implementing ActivityPub is notoriously difficult due to things like JSON-LD, and it pitches a library called Fedify to fix it. Fedify, for the uninitiated, is a framework designed to make building ActivityPub servers less of a headache. Samantha, is this a solved problem now?

GuestI wish it were that simple. The author makes a solid case, but you can feel the tension in the Lobsters thread. One user, morj, makes a great point that developers usually end up just forking existing projects because figuring out someone else's mess is often faster than dealing with the protocol's inherent weirdness. I honestly felt that. Using JSON-LD as a real linked document is just a recipe for pain. You end up dying by a million corner cases, just like another user, mdaniel, pointed out in the thread. It is a classic case of a library trying to be the hero, but I do not think the community is sold on the idea that an opinionated wrapper is the only way forward.

HostIt sounds like the "it works on my machine" version of federated social networking. Moving on, we have a fascinating post about the internals of LLM tool calling. The author claims that newer models like Opus 4.8 are actually getting worse at following strict tool-calling schemas because they are being over-optimized on specific agentic harnesses. They end up inventing extra keys that cause everything to break. Samantha, have you hit this wall?

GuestOh, absolutely. This is the "agent hallucination" tax we are all paying right now. When the model thinks it is smarter than your schema, you are in for a long afternoon. The author is right, this feels like these models are being reinforced on messy, agentic data, and now they just cannot help but freestyle. I actually started using a custom validator in my own dev environment that strips out any extra parameters before they even touch the API, just to keep the models on a short leash. It is not the most elegant fix, but it saves me from watching my pipeline crash for the tenth time in an hour. One commenter, Wilfred, even mentioned that the classifiers are getting aggressive enough that simple tasks get flagged as security research. It is a weird time to be a builder, Daniel.

HostIt is a delicate balance. Finally, let us look at the engineering side of this, specifically GitHub Copilot in VS Code. The team behind it wrote a breakdown of how they are squeezing more out of their token usage with prompt caching and deferred tool loading. They are basically making the argument that since we have usage-based billing now, efficiency is no longer optional.

GuestThis is the kind of stuff I actually love. It is not glamorous, but it is necessary work. The chart they showed of tokens per turn increasing over time? That is the hidden cost of "intelligence" that no one wants to talk about. The VS Code team is doing the grunt work here with A/B experiments and offline evals. It is a grind, but those small gains in cache hit rates are what make a tool feel snappy instead of sluggish. It is the difference between waiting for a thought and waiting for a server farm to finish calculating.

HostIt sounds like the era of "just throw more context at it" is officially over. Samantha, it has been a wild ride through these threads today. Any final thoughts before we wrap?

GuestJust that I am definitely going to be obsessing over my own token hit rates for the next week because of that VS Code post. And honestly? I am glad someone is finally documenting the pain of JSON-LD so I do not have to feel alone in my suffering.

HostThat is the beauty of a community like Lobsters. Thank you for joining me, Samantha, and thank you all for listening. We will be back tomorrow with more.