Craft & Chaos: Episode 2 Highlights

Ryan and Dario discuss control over conversation state, iterate on a Java desktop animation, and demonstrate skills and an experimental interactive-tool loop. This summary uses the official YouTube recording's English auto-generated captions, not an independent audio transcription. Feature descriptions reflect the December 2025 broadcast, not current Amp behavior.

• Convenient state management can constrain later changes to an agent. Discussing Google's Interactions API, Dario contrasts clients that keep and resend conversation history with a provider-managed conversation. Keeping state in the client takes more work but lets the developer filter or transform it before an inference call. His concern is losing that control when relying on server-managed state; the discussion is an architectural preference, not an exhaustive account of the API's capabilities. 12:34–14:31(opens in a new tab)
• Visual iteration needs both a running result and compiler feedback. Dario asks Amp to make a Java desktop animation resemble liquid emerging from an orb, then adds word-by-word text and a viewer-requested neutron star. The first star implementation does not compile, so he runs the command in Amp's shell mode to expose the error output directly. They then adjust the rendered animation's speed. The sequence shows exploratory work with disposable code, not a claim that the first generated implementation works. 18:20–27:14(opens in a new tab)
• A thread reference carries selected context, not the entire old conversation. Continuing the animation in a new thread, Dario explains that another model extracts information relevant to the new request from the referenced thread. That result enters the main agent's context. He presents this as his alternative to handoff; the demo does not establish a fixed compression ratio or guarantee that the next visual change will be correct. 27:14–30:30(opens in a new tab)
• The skills prompt duel fails as a controlled comparison. Ryan asks one agent for homepage-copy feedback and explicitly asks the other to use a copywriting skill. The first agent discovers the skill anyway, while the second starts applying changes Ryan did not want. Dario explains the useful mechanism behind skills: the agent knows instructions are available and loads them when needed, rather than carrying every instruction in AGENTS.md all the time. The demo illustrates automatic discovery and differing actions, not proof that one prompt produces better copy. 32:21–38:08(opens in a new tab)
• Instructions and tools serve different roles. Dario describes tools as precise operations, whereas skill instructions still require model interpretation. He treats MCP as a convenient way to distribute tools such as Chrome DevTools, not something skills inherently make obsolete. Loading relevant tools alongside a skill is an idea he would like to see, not a demonstrated feature in this episode. 41:08–46:07(opens in a new tab)
• An experimental tool connects a model to an interactive process. Dario demonstrates changing a running Node.js request handler through node inspect; the initial attempt needs debugging before the new message appears without stopping the server. He explains the loop: model output becomes subprocess input, subprocess output returns to the model, and a stop tool ends the session and reports back. A second demo uses psql schema-inspection commands to find a thread in his local development database. The feature is explicitly unreleased in the broadcast. 48:06–55:17(opens in a new tab)

Recording: Watch on YouTube(opens in a new tab) · Original X post(opens in a new tab)

Build Crew: https://buildcrew.team(opens in a new tab)