#buildinpublic
Yesterday I came across Haptiker (https://t.co/rBxbgYolzT)
on Product Hunt. It lets you adjust your Mac’s volume by sliding along the trackpad edge.
Thought it would be fun to build my own, so I made EdgeVolume.
Still a work in progress, but yeah, here is the repo
https://t.co/0VBZCbdUVD
#opensource
If you give an AI agent direct permission to execute third-party API mutations, you are begging for duplicate charges and quota exhaustion.
LLM retries and tool-calling loops will eventually fire destructive POST requests multiple times.
Here is the architecture pattern top agentic teams use instead:
The Two-Phase Shadow Tool Gate
The Flow:
1. Split agent tools into two tiers: Read-Only (Unrestricted) vs Mutating (Gated).
2. When the agent calls a mutating action (e.g., charge_card or cancel_subscription), it cannot execute directly.
3. The proxy generates an immutable execution manifest containing:
- Target endpoint + payload
- Deterministic Idempotency Key (agent_session:step_id)
- Expected state diff
A human or deterministic policy engine signs the manifest before the payload touches production network sockets.
The LLM decides intent; a deterministic state machine handles execution.
Trying to make my feed feel less like a timeline and more like a room full of interesting people.
If you’re building a product, writing code, running a business, designing, or simply figuring things out say hi and tell me what you’re working on.
I’d love to follow your journey.
#connect
@heyclairelin Fair. A useful filter is asking people to reply publicly with why they want to talk before sending anything private. Genuine people can give context; scammers usually want the conversation hidden immediately.
@karthiknish Probably people whose work already starts in X DMs—creators, founders, recruiters. The separate domain makes it feel like X wants calls to become a standalone workspace, not just another button in messages.
@elshayib_ The code may become cheaper and more open. Reliable operation probably won't: hosting, migrations, support, compliance, and knowing who fixes production at 2 AM still have value.
@Declan_on_X Hey Declan! I'm on the developer side, building tools for store operations and document-heavy work. What part of business or tech are you focused on?
@mikedutuandu Hey Mike, let's connect. Curious about SiteSeed: after the site scan, does it help implement the fixes or give the founder a list to work through? That handoff seems like a big part of the value.
@ayoub_el77927 Hey Ayoub, let's connect! I'm building InfiBis for store operations and an AI tool for engineering documents. What's the project getting most of your attention right now?
@Anmol1578 Nice, Anmol. A useful next challenge: open a specific screen from a link while logged out, then land there after login. That gives all those navigation pieces a proper workout.
@dotnetsme An interrupted spreadsheet import is a good test. Does it leave half the ledger posted, or roll everything back? That's the sort of behaviour I'd want defined before generating the implementation.
Good morning everyone.
My morning started with
AI is making me look at Rust differently.
Not because it replaces Python, but because someone still has to build the infrastructure that makes AI fast, reliable, and scalable.
Python builds a lot of the AI.
Rust can help make the systems around it faster and safer.
I'm a MERN dev moving into AI, and I think it's time to stop building just another chatbot and start understanding what's happening under the hood.
Time to learn Rust.
Let's stay #connect along the journey