Diego Sapriza (@AV4TAr) header
Diego Sapriza (@AV4TAr) profile picture

Husband and father of 4. SVP Engineering @playbypoint, former XOPS, @WeWork, @case_inc Optimizing for collaboration.

CA, USA
Joined November 2008
Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)Photo by Diego Sapriza (@AV4TAr)
100% de acuerdo. Y creo que se ha expandido no solo a verticalmente (depth) pero horizontalmente mucho más (breadth) Ahora el trabajo es más end-to-end y eso lo hace complejo.
No conozco a nadie que haga uso intensivo de la IA y trabaje menos que antes. La posibilidad de hacer cosas en paralelo te permite ir más rápido, pero te obliga a pensar en varias cosas a la vez. Además, una vez metido en la dinámica, cada rato “libre” se siente como improductivo, lo que hace que estemos constantemente pensando en el siguiente paso para no perder el ritmo.
40
Diego Sapriza retweeted
57
517
52
7,992
788,794
Diego Sapriza retweeted
Anthropic just published an engineering guide on how to build always-on agents. The biggest misconception: an always-on agent does NOT run 24/7. It sleeps almost all day, wakes up for short runs, saves state, and goes back to sleep. The 5 core pieces I distilled: 1. Event Inbox Put a database table between webhooks and your agent. Acknowledge webhooks fast, run only one active worker per user so runs never collide, and batch rapid messages together. 2. Timers Keep timers in a separate table from the inbox. Split them into one-off follow-ups and recurring schedules. A background poller drops due timers into the inbox. 3. Lean Memory + Dreaming Only store what cannot be looked up again: user instructions, user notes, past messages, and task progress. Load a small index at wakeup, not every file. Run periodic dreaming passes to clean out stale facts. 4. Unattended Permissions Never put raw API keys inside the agent sandbox — inject them through an outside proxy. Use Deferred Tool Calls to pause risky actions for approval without blocking the worker. 5. Outbox Queue outgoing replies so your agent paces notifications instead of spamming users at 3 AM. claude.com/resources/article…
76
144
17
3,117
245,214
Diego Sapriza retweeted
Me explaining how my app works after vibe coding the entire thing.
482
1,624
261
20,411
1,371,090
What!!!!!
Unfuck your hardware with Hermes
64
Diego Sapriza retweeted
Open source maintainers… hope this helps
It's a scary time to be an open source maintainer. Everything that can be exploited will be exploited. If you aren't using abliterated Chinese models and frontier cyber models on your code...someone else is. Sharing my workflow for red-teaming XMTP. gist.github.com/neekolas/99a…
1
6
936
Every methodology eventually collapses into this one. The rest are just what we call it in the retro.
39
674
51
4,632
99,416
Diego Sapriza retweeted
Me casually dropping a 3,000-line PR twenty minutes after the ticket was assigned.
25
145
8
3,664
140,201
Diego Sapriza retweeted
This is why MCP mogs an API. An agent could send an email to 1000s of people via an API, without you being able to stop it. With MCP elicitation, the agent initiates the call, but NEVER touches the confirmation. It goes straight from the MCP server -> client & you click send. In this example, imagine we tell an agent to send an email to all 2k of our subscribers but notice some info is wrong. We want to cancel it and fix that before it goes out. This is a "destructive" action that a overly proactive agent could do on its own via an API. Even if you sent it a deeplink back to your dashboard it could use computer use or devtools MCP to click the confirm button. With MCP, the agent never sees the confirmation button, it just goes straight to your MCP client (ChatGPT, Claude) and you have to confirm.
18
9
2
122
20,801
Amazing news!!!
As reported in the @WSJ, we have raised a Series B to bring Hermes Agent to new frontiers (and, yes, build a mobile app) Thank you to our investors including @nvidia @M12vc @SamsungNext @robotventures @usv @ycombinator @MenloVentures, among others wsj.com/pro/venture-capital/…
1
135
Diego Sapriza retweeted
.@OpenRouter co-founder Alex Atallah on the "everyone is building the same thing" take: "This reminds me of this tweet I saw... Everybody's building an agent loop with notifications, third-party connectors, context management, memory, sandboxes, agentic web search, an always-on agent on top." "This product is showing up everywhere. It is showing up everywhere, but these are just the new table-stakes primitives." "A 2005 version of that tweet would be: 'Oh, everybody's building the same thing. A database, a users table, a sign-in page, a sign-up page, a profile page, a logout page. Everything's the same.'" "There's like a lot of differentiation really. There's table-stakes needs for AI just like there's table-stakes needs for the web." @alexatallah @amasad
.@OpenRouter co-founder Alex Atallah, in his first podcast since Stripe acquired the company, joins @Replit co-founder Amjad Masad and a16z's Erik Torenberg on why the future of AI is independence and specialization. In this conversation, Alex walks through how the Stripe deal unfolded, why he wasn't originally looking to sell, and why "payments and inference are going to blend together." Pre-OpenRouter, the typical AI workflow had one model provider to choose from, and little pressure on that provider to lower prices. Now enterprises are diversifying across labs and open-weight models, and every board is asking about AI costs and benchmarks. Amjad argues if your company depends on one AI lab, it can turn into your competitor. So Replit is building the layer that lets enterprises use any model and any cloud, without being locked into either. Alex and Amjad are split on personal agents – Amjad runs one agent across his whole company and loves the cross-domain joins, while Alex says general agents cause you to sacrifice understanding, and argues 10 specialized chiefs of staff beats one superagent. 0:45 How the Stripe deal unfolded 5:05 Why mixing models beats one model 7:25 Forcing the labs to compete on price 8:50 Enterprises want open-weight models 10:30 Every board asks about AI every month 12:25 Why companies must own their intelligence 14:15 Replit as the independence layer 15:10 Everyone is building the same agent 16:35 Why Amjad built bring-your-own-cloud 18:10 Amjad's agent that runs his whole company 19:55 Why 10 specialized agents beat one 23:35 Machines, not humans, should specialize 27:30 Guardrails for agents talking to agents 31:10 Models training their own replacements 33:45 Most tasks don't need a frontier model 40:50 Training small models on Qwen 8B 43:25 The Rust cycle is coming for AI 45:10 Fusion models: frontier quality at half the cost YouTube: youtu.be/ekK8urKHPMQ @alexatallah @OpenRouter @amasad @eriktorenberg
148
288
106
4,188
724,730
Diego Sapriza retweeted
You can now mod Claude Code: - Change how it behaves - Customize the UI - Swap in your own features Write one with a few lines of TypeScript, or have Claude build it for you. Mods ship inside plugins, so you install them with /plugin in the CLI or desktop app. A few examples:
698
1,235
816
20,346
4,547,729
Diego Sapriza retweeted
Opus 5.5 with a 1M window in Hermes Agent is brilliant, until every turn re-reads 800K tokens of old chat and your Claude limit is gone by lunch. Here's the compression setup I landed on. It keeps Opus sharp and takes the summary work off your Claude plan completely. What I learned: - The 1M window costs nothing by itself. What burns allowance is how much context each turn re-reads. Compress early and Opus stays fast and focused. - Set the trigger per model, not globally. Opus compresses at 25% (about 250K). GPT models at 85%: about 231K on the 272K route, about 765K on the -900k models. - The summariser doesn't have to be Claude. I use GPT-6 Luna on high through my ChatGPT plan, so compression never touches my Claude limit. - Use gpt-6-luna-900k, never plain gpt-6-luna. Hermes won't let a summariser read more than its own window, so plain Luna's 272K quietly pulls every bigger trigger down to 272K. - Keep lean mode. One summary call, your messages kept word for word, and anything summarised away stays searchable. No ChatGPT plan? Put Sonnet 5 [1m] in as the summariser instead. Same idea, it just comes out of your Claude allowance. Copy and paste into ~/.hermes/config.yaml: compression: enabled: true threshold: 0.50 threshold_tokens: null tail_mode: lean in_place: true model_thresholds: "claude-opus": 0.25 "gpt-6": 0.85 "gpt-5.6": 0.85 auxiliary: compression: provider: openai-codex model: gpt-6-luna-900k reasoning_effort: high timeout: 300 Why Hermes has to own this: the Claude subscription plugin turns off Claude Code's own auto-compact, so Hermes' trigger is the only compaction that runs. Docs and the plugin: Docs: Hermes compression: hermes-agent.nousresearch.co… Hermes config: hermes-agent.nousresearch.co… Claude subscription plugin: github.com/NousResearch/herm… Claude Code auto-compact window: code.claude.com/docs/en/mode… Claude Code costs: code.claude.com/docs/en/cost…
25
19
6
382
49,043
Diego Sapriza retweeted
This has never been more true. The opportunity cost of endless planning and contemplation just went way up. Agents allow you to just try vastly more stuff, so let them, and you'll learn way more, way faster.
This quoted post is unavailable.
126
457
34
5,296
256,387
Diego Sapriza retweeted
156
483
95
7,045
863,627
Diego Sapriza retweeted
I'm looking for the next batch of 5 users to test the new @teal_hq. Here are some criteria to decide if you're a fit: - You're looking to make a career move (in the US). - You're NOT focused on your resume. - You are a networker, or want to network better. - You must be willing to engaging our private Discord. - You'll provide feedback. - Active AI user (Claude, ChatGPT, MCPs, Skills, etc) - Google (Gmail) and iPhone user - 1 hour onboarding zoom. Given privacy needs around career moves, feel free to DM me.
4
4
12
828
WOW
My holy-shi moment of opus5. Homeworld Style Space RTS with OPUS 5 (workflow) 68.6k input, 4.6m output, 837.2m cache read, 13.6m cache write ($632.65) --- Models: generated, Asset: none. Besides minor game play issues, watching model cook this is beyond my imagination.
5
1,194
Diego Sapriza retweeted
Elon Musk literally broke down his 5-step process for applying first-principles thinking to build anything:
33
486
10
2,868
599,585
Well, this got weird: When you need the user's attention (blocked, async task done, need input), use: say -v Whisper "<short message>" Keep it a nudge — one short sentence max.
1
52