Juan Esteban Sierra

Partner at - Frontier Tech Growth Partner🧙‍♂️🔥
Board @2U2ai
Shaping Web3 Through Culture, Strategy, and Story
@_buildspace s4 gaudmire

[Juan Esteban Sierra @juanesbuilds]

@nextbigfuture Speeding up one step just moves the queue to the next one. Code generation gets faster, then review becomes the bottleneck, then QA, then handoff. The 10X only shows up when those get redesigned around the new speed, not when the first step gets quicker.

[Juan Esteban Sierra @juanesbuilds]

@KostyaAI Curious how it handles version drift: does the agent pin the docs it retrieves to the SDK version in the repo, or does it pull whatever is current? Current docs against an older lockfile is its own kind of stale.

[Toxic Cowboy 🤠 @toxiccowboy02]

One of the best I've seen so far 😜

[nicksaraev @nicksaraev]

I am not joking when I say this.

Traditional game development is over.

I asked Opus-5.5 to make Attack on Titan in Minecraft. Took ~1hr and cost less than $25 start to finish.

The end result is extremely fun. And multiplayer!

Guide + link to play below.

[Beaver @Beaver_0x]

WHOEVER LEAKED THIS CLAUDE CODE SETUP DESERVES A RAISE

I thought the best way to use Claude Code was to just throw Opus at everything.

Apparently not.

Anthropic’s setup is much more interesting:

→ Sonnet 5.5 runs the main session and actually builds the thing
→ Haiku 5.5 swarms the repo for file discovery, AST scans, grep, and docs
→ Opus 5.5 stays in the background as the architect
→ Opus gets called in before big plans, after repeated test failures, and before shipping
→ JEV handles the boring mechanical routing in ~16ms so frontier models only wake up when real reasoning is needed

so you get:

Opus reviews. Sonnet builds. Haiku swarms. JEV routes.

the interesting part isn’t using multiple models.
it’s giving each model a different job.

Instead of burning Opus context on finding files and running routine commands, you save it for the moments where architectural judgment actually matters.

Plan on high.Delegate on medium. Keep Opus on call.

[Juan Esteban Sierra @juanesbuilds]

@robertdoneill Is that version token per cell or per sheet? If it is per sheet, two agents writing disjoint cells will keep refusing each other, and the retry loop becomes the bottleneck.

[Juan Esteban Sierra @juanesbuilds]

@ObedOnyedika1 The cheap gate for the terraform case is making the agent apply only a saved plan file, and rejecting it if terraform show -json lists any resource_changes with a delete or replace action. The volume replacement is visible in the plan before anything runs.

[Juan Esteban Sierra @juanesbuilds]

@NickAbraham12 You give 70 to 80 percent ICP accuracy for the old Apollo pull. What does the AI-validated list score when someone spot checks it by hand? Curious whether the validation step has its own error rate nobody measures.

[Juan Esteban Sierra @juanesbuilds]

@NatureBlueee Write side. A uniqueness check on source, target and type inside the gate beats any process check, because the check you cannot trust is the one that fails the day it matters. We key retries on the work item, not the process, so a twin's second write just bounces.

[Juan Esteban Sierra @juanesbuilds]

@VentureBeat The approval that works is the one asked before the action runs. Once the agent has already shipped, the sign-off is just a notification with a question mark.

[Juan Esteban Sierra @juanesbuilds]

@samtoshi7 Proof recorded on-chain is an audit trail, not a control. What decides if this works in production is whether the policy check blocks the call before the agent touches the system, and who updates the allowed set when the agent's tools change every week.

[Juan Esteban Sierra @juanesbuilds]

@robertdoneill Worth noting Sheets values.update has no precondition on the version you read, so the stale write always wins. The fix has to live in the harness: re-read the cell right before writing and refuse if it changed since the agent's read.

[Juan Esteban Sierra @juanesbuilds]

@FundamentEdge The hard part we keep hitting isn't the one-shot, it's the harness: validating that a generated number actually traces back to the filing it cites. Most teams skip that step until a wrong figure gets repeated three tabs downstream.

[Juan Esteban Sierra @juanesbuilds]

@marcusyul Eso es justo el cuello de botella real del multiagente: no es el modelo, es el traspaso. Cada vez que un contexto se copia a mano entre dos agentes se pierde algo, y nadie lo nota hasta que el output sale raro tres pasos despues.

[Juan Esteban Sierra @juanesbuilds]

@github Matches what we see running agent pipelines for clients: co-authored-by trailers are opt-in, and most harnesses skip them unless you wire it in yourself. So the undercount is probably the norm, not the exception.

[Juan Esteban Sierra @juanesbuilds]

@ash_twtz In practice the gap doesn't close, it moves: you stop writing the function and start writing the constraints the agent gets wrong. What's left for engineers is spotting the output that looks right and isn't, which is a harder read than writing the code was.

[poet.base.eth @1CrypticPoet]

just got a call from my kid's teacher.

[Juan Esteban Sierra @juanesbuilds]

@moonpay

[Juan Esteban Sierra @juanesbuilds]

@OSHBuilt The harder part isn't standardizing the MCP servers, it's liability. Once an agent picks a shop and specs a part off DFM data alone, someone has to be on the hook when it fails a tolerance check. That gates adoption faster than the protocol maturing does.

[Juan Esteban Sierra @juanesbuilds]

@mfishbein The failure mode we've hit with this: a rubric calibrated to one reviewer's taste overfits, and the next reviewer disagrees with the score, not just the thumbnail. Log inter-reviewer disagreement before trusting an 8.8 to 1.2 drop as signal.