1. 5 concrete, fixable faults in @Nick Vasilescu template — Been running @Nick Vasilescu's template for a a bit over a week now and want to lay out what's actually broken in it, because I think the honest list is more useful to this group than pretending it's plug-and-play. 2. The infra faults: the root-vs-orgo user split is the single biggest time sink by far. There's an editable-install permissions trap. Hermes has a real CLI bug around plugin discovery. Ad-hoc restarts fight Orgo's own supervisor instead of working with it. 3. Then the integrations, which is where it gets worse. AgentCard's OAuth callback assumes a local browser, and even once you work around that, we hit a genuine Orgo browser-backend outage on top of it, two separate failure points stacked on one integration. AgentPhone quietly needs 10DLC carrier registration before outbound SMS works at all, nothing tells you that going in. AgentMail wants two separate credentials, not one, easy to half-configure without noticing anything's wrong. And "no credits monitoring" turned out to be the small version of a bigger problem, a Composio auth token expired on us silently this week, no alert, nothing. Credits and credentials fail differently and this template only thinks about one of them. 4. Here's where I land on it though: none of this is a reason not to sell it. A business owner would have zero path through a root-permissions bug or an OAuth callback fighting a cloud browser outage. That gap, between "looks like a 30 minute video" and "is a multi-hour infra debug across four integrations," is the actual product. Just don't sell it as a self-serve course, because it isn't one, even with direct root access. Even our own watchdog needed two more real fixes after we called it "done." Sell the done-for-you build with an ongoing retainer attached, because this stuff doesn't stay fixed on its own. 5. Genuinely want to know if others are hitting the same wall or if I'm missing a cleaner path, especially on the AgentCard/browser side. Tell me where I've got this wrong.

Hi @Nick Vasilescu Token spend reality check on a Chief-of-Staff style agent setup I've been running an HERMES agent on Orgo cloud desktop for a few weeks now: Chief-of-Staff style setup, driven through a messaging app, delegating to sub-agents for a single project. Token spend has hit a point I genuinely can't sustain, and I want to check whether I'm missing something obvious or this is a known pattern others have hit. The numbers, last 7 days: - 51.4M total tokens, ~12.6M input / ~197K output, across 8 sessions - Messaging-app sessions: 46.5M tokens. CLI sessions: 4.9M tokens. Same agent, same work, wildly different cost by platform - Average session: ~1.2 days and 70 messages before getting reset My working theory (correct me if wrong): messaging-app sessions don't get reset often, so every new message resends the entire accumulated conversation history to the model. A session that's grown to 300+ messages costs many times more per turn than a fresh one, even for a trivial question. Not a bug, just how context windows work, but it means casual "always-on chat" usage is fundamentally more expensive than deliberate CLI sessions unless you're disciplined about resetting. Questions for the community: 1. Is aggressive session-reset discipline really the main lever here, or is there a better pattern for an "always-on chat assistant" that doesn't force a full-history resend every turn? 2. Anyone running multi-agent/delegation setups (sub-agents, skills, orchestration) who's found a way to keep the orchestrator session lean while delegating the token-heavy work elsewhere? 3. Is there a standard "cheap model for routine stuff, expensive model only when needed" pattern people use in production, and how do you actually enforce it so it doesn't get ignored in the moment? Not blaming the tooling, just trying to figure out if I'm missing a known best practice before this becomes unaffordable for a solo operator.