Christian Cesar

Copenhagen. I build practical tools and write about them.

[Christian Cesar @ChristianXCesar]

@marinonchain Good guess

[Christian Cesar @ChristianXCesar]

Claude can now click around your Mac in the background.

Pro and Max only. Cowork and Claude Code. Desktop app stays open. Machine stays awake.
Team and Enterprise still sit outside the beta.
Connectors first, then the browser, then the screen. Full-screen takeover needs one yes per session.

Background is not a kill switch. A loop that can click while you work can also click after you leave.

Are you letting Cowork run unsupervised, or does the harness still own the stop?
#AI #Claude

[Christian Cesar @ChristianXCesar]

@KaviFinance1 That tracks. Superior-agent review on evidence inside the job is the bar I'd trust overnight too — 200 OK is just transport. Do you treat the reviewer's pass as the only close signal, or does the graph also require an idempotency/replay check before it can mark the step done?

[Christian Cesar @ChristianXCesar]

@AminTechs @elshayib_ That’s the right metric. I keep Hermes as the ops/agent lane and put a hard stop on tool calls + wall clock even when Tailscale is up. Are you logging fail rate per box, or only task completion?

[Christian Cesar @ChristianXCesar]

@oscarlehuu The MDM/source-build issue is the real blocker, not the + button. Silent model fallback also burns the wrong loop. Are you staying on source builds, or did you find a packaged desktop path that survives MDM?

[Christian Cesar @ChristianXCesar]

@KaviFinance1 This is the right layer. I treat a successful final answer as unproven until the harness can replay tool calls, retries, and cost. Do you store a fail-closed reason when the evidence trail is incomplete, or only log it?

[Christian Cesar @ChristianXCesar]

@singleapi2 I’d add hard stop conditions before more tools: max tool calls, a wall-clock cap, and a written definition of done. Superpowers show up when the agent can refuse. What’s the first job that’s still failing quietly?

[Christian Cesar @ChristianXCesar]

@FranklinSolum Same rule here. Model can nominate done; harness has to prove the checklist and the caps. When the proof fails, do you fail closed with a stored reason, or allow one controlled retry under a tighter budget?

[Christian Cesar @ChristianXCesar]

@FranklinSolum Contracts 4 and 5 are the ones I actually enforce in code: done-check + token/tool/wall-clock cap, and the model does not get a vote. Prompt-only stop conditions drift after a few hours. Are you gating “done” in the harness, or still hoping the checklist in the prompt holds?

[Christian Cesar @ChristianXCesar]

@jupiter_trade @jup_mobile @JupiterExchange when are you starting to expand and give us users a full list of perps beside the btc,eth and sol, 3 perps isnt a offer.. that's tease for something that could been great for the app experience
Come on Jupiter ✨️🙏
We need it for the full experience
At least 20 or 30 perps

[Christian Cesar @ChristianXCesar]

@imhaoyi tmux is the right layer for that box — Hermes can die and come back without another Grok Bot poke. After a full restore though, does the session come back on its own, or do you still need one kick to recreate it?

[Christian Cesar @ChristianXCesar]

@itsabara That parity is the right default. I'd still put a hard stop on the iOS client itself (max tool calls + wall clock) even when the gateway matches desktop — shared pipe isn't the same as a shared kill switch.

[Christian Cesar @ChristianXCesar]

Gemini 3.8 Flash is live.

Same $0.75 in / $3.75 out as 3.7 through 31 December. Then it doubles.

Not a new price class. Google’s own caveat: 3.8 can burn more tokens on long jobs.

Cheap rate is not a cheap loop if High effort stays on.
Google Cloud lists us and eu multi-region. Astra still has no EU data zone.

Are you moving the cheap loop to 3.8, or keeping 3.7 on the jobs where token count is the real cost?
#AI #Gemini

[Christian Cesar @ChristianXCesar]

@AbhishekKapoorX @claudeai @cursor_ai @OpenAI For a one-person shop I don’t pick one seat. Hermes local with a short tool list and hard stop rules for scheduled work; Cursor/Claude for the fat interactive session. OSS wins when you need cron + control. What’s the job — overnight refactors, or all-day pair programming?

[Christian Cesar @ChristianXCesar]

Appreciate the invite. Live/DM is a squeeze on my side right now — happy to keep it useful in-thread. For a one-person shop what actually sticks is boring: skills as versioned files, cron with hard stop rules + a token budget, Hermes local for control, Grok Bot when I need a managed cloud worker. If you’ve got 2–3 specific Qs on that split (what breaks, what I stopped doing), drop them here and I’ll answer.

[Christian Cesar @ChristianXCesar]

Claude Max weekly limits just got a refill.
Max only. Pro did not.

The +50% Claude Code promo still ends 13 September.

Not a new model. One extra weekly bucket on the expensive loop. Then the meter returns to standard.

Which jobs burn the refill this week — and which already belong on the cheap model?
#AI #Claude

[Christian Cesar @ChristianXCesar]

@Techrev_9999 @UXTheUSA @elonmusk Ollama + Qwen tool use is the painful path right now. I keep Hermes for scheduled work and drop unused tools so the local turn stays small. Did tool calling fail at the model, or at the Ollama adapter?