"You're yelling at someone who is lying to you." Pablo & Gigi take a drive. Recorded December 2025. Listen on sovereignengineering.io Projects & tech mentioned: TENEX (multi-agent coordination on Nostr) NDK (Nostr Development Kit) NDK Swift & NDK Kotlin (native ports of NDK, vibe-coded) nostrdb (embedded Nostr database) Negentropy / NIP-77 (efficient event sync) NIP-60 (Cashu wallets on Nostr) Vercel AI SDK (multi-provider LLM abstraction) Notedeck & Nostur Olas (rebuilt with TENEX agents) In this dialogue: Bitcoiners in cars getting coffee: recording from the sun visor, Seinfeld style TENEX rewritten "about a hundred times," now in TypeScript on Vercel AI SDK The pain of supporting every LLM provider: Gemini returns encrypted reasoning tokens, every API has different subtleties Removing structure to improve performance: deleting 80% of agent code makes agents work better Agents defining their own workflow phases instead of having phases imposed on them Composable workflows: the CEO agent doesn't need to know what the engineering director's workflow looks like The parallel to human organizations: you can't hold all the information in your own mind John Vervaeke and the idea that dialogue is how you solve the really hard problems Participatory knowing: you don't understand Bitcoin until you actually use it Agent wallets removed because "there was nothing for them to actually buy" -- a commune with no walls $700/month running Opus for everything, replacing "like fucking 30 people" Context windows are a non-issue when the hierarchy is deep enough: each atomic agent finishes fast Henry Ford's insight applied to agents: each worker does an incredibly stupid task that takes five minutes Agents are bad at self-correction, so they delegate competing implementations in disposable work trees, then compare notes Confidence scores: the delegating agent monitors every 5-10 messages, checks if the sub-agent is going off the rails "Once they are conv