Agent infrastructure requires runtime policy enforcement, not just model-level guardrails. Deterministic controls at the system boundary prevent policy violations regardless of model behavior
Check the post: https://www.supra-wall.com/blog/llm-as-judge-fails-agent-sec...
Yes, because we've built them to be flexible. The fix requires treating agents like any other service: least privilege, API quotas, audit logs, separate validation layer. Not guardrails—governance.
Exactly. Today's agents inherit LLM semantics (ambiguous, probabilistic, interpretable). Real solution: agents operate in ACL model (access control lists) tied to execution user, not prompt understanding. https://www.supra-wall.com/en/blog/llm-as-judge-fails-agent-...
Hey HN! I built AgentGate, an open-source security firewall for AI agents built on OpenClaw.
The problem: AI agents (browser, bash, fetch) can call any tool without restriction. There's no built-in way to define policies like "agents can read files but not delete them" or "require human approval before sending emails."
AgentGate solves this with a policy evaluation layer that wraps your OpenClaw tool calls and enforces ALLOW/DENY/REQUIRE_APPROVAL rules in real time.
How it works:
1. You define policies in Firebase Firestore (regex-matched against tool name + args)
2. Every tool call hits the AgentGate middleware before execution
3. ALLOW → passes through, DENY → blocked immediately, REQUIRE_APPROVAL → Telegram webhook fires, human approves/rejects in real time
4. Full audit log stored in Firestore with timestamps, agent ID, tool called, decision, and approver
Stack: Next.js dashboard, Firebase Firestore + Auth, OpenClaw bash/browser/fetch agents, Telegram Bot API for approval notifications.
What's working in v1.0.0:
- Policy engine (ALLOW/DENY/REQUIRE_APPROVAL)
- Real-time audit log dashboard
- AI Policy Wizard (describe a rule in English → generates the regex policy)
- Telegram approval flow
- Tested against OpenClaw bash, browser, and fetch tool types
Would love feedback on the policy model, the approval UX, and whether this is useful for teams running autonomous agents in production.
Stop wasting ad spend on invisible products. Magnify is the first AI tool designed specifically to fix 'Title Truncation' on Google Shopping mobile ads.
The Problem: 60% of Google Shopping traffic is mobile, where titles are cut off after 70 characters. If your key attributes (Size, Color, Model) are buried at the end, users scroll past your ad.
The Solution: Magnify analyzes your entire product feed and automatically rewrites titles to 'front-load' high-intent keywords.
Key Features:
Visual SERP Simulator: Preview exactly how your ads look on iPhone vs Desktop before you spend money.
Bulk AI Rewriter: Optimize 10,000+ SKUs in minutes, not weeks.
Supplemental Feed Sync: We never touch your original Shopify/WooCommerce data. We create a safe, reversible layer on top via Google Merchant Center API.
A/B Testing: Automatically test new titles against old ones to prove ROI.
Trusted by ecommerce managers to boost CTR by 25% or more without increasing ad budgets.
I built Prompt2Sheets because I kept watching non-technical founders and operators lose evenings inside Google Sheets – fighting VLOOKUPs, pivot tables, dedupe formulas, and ad/lead exports that never line up. They didn’t want another BI tool; they just wanted their existing sheets to behave.
Prompt2Sheets lets you edit a Google Sheet by describing what you want in plain English. Under the hood it generates JavaScript, runs it in a sandbox, and syncs the result back to the original sheet. Typical prompts are:
– “Remove duplicate leads by email and keep the most recent row”
– “Add a Margin column after cost of goods sold = (price_fyi – cost of goods sold) / price_fyi”
– “Filter to only US + Canada with verified emails and add a summary row with total spend”
It’s aimed at people who live in Sheets (founders, operators, recruiters, marketers, ecommerce) but don’t want to learn complex formulas or maintain brittle scripts. The focus right now is: cleaning lead lists, transforming ad/export data, adding calculated columns, and building simple summaries directly in the sheet.
You can try it here: https://prompt2sheets.com
– No credit card required
– Uses Google OAuth; app is in beta but live
– Works best on small/medium sheets while I harden performance and edge cases
I’m particularly interested in feedback on:
– Where the natural-language interface feels useful vs. confusing
– Which operations you’d trust it with (and which you would never hand to an AI)
– Any cases where it silently does the “wrong” thing – those are the most valuable to fix
If you have a messy sheet you’re willing to share (leads, ads, inventory, reporting), I’m happy to personally help set up your first workflow and debug failures. Honest criticism is welcome – especially around UX, safety, and where this is obviously the wrong tool.
I just launched AskGAAI, an AI analyst for GA4 + Google Ads built around a very SaaS‑founder problem:
We had all the data… but still couldn’t answer basic growth questions:
Why did signups drop this week?
Which campaigns are actually profitable, not just cheap clicks?
Where exactly are users leaking in the funnel?
What should we fix first to move MRR instead of random metrics?
I was spending hours inside GA4 and still guessing. By the time we noticed issues, revenue or ROAS was already hurting.
So I built AskGAAI to watch GA4 + Google Ads and tell you, in plain English:
1⃣ What changed – traffic, signups, CAC, key events
2⃣ Why it changed – which channel, campaign, or step in the funnel moved
3⃣ What to do next – concrete recommendations, not just another chart
On top of that, it:
Sends a daily email brief with your key metrics, 3 insights, and 1 recommended action
Triggers real‑time alerts when signups or conversions drop, CAC spikes, or tracking breaks
Lets you ask questions like a human (“Why did trial → paid conversion fall yesterday?”) instead of clicking 10 GA4 reports
Remote hiring across time zones has turned my calendar into a war zone.
For one backend role, I ended up doing ~20 interviews across 6 time zones. Early mornings, late nights, constant rescheduling, no-shows – and still felt like I was guessing at the end.
I built EasyHire to offload that entire first round.
What it does:
You drop in a JD or a batch of CVs.
Candidates get an async AI-led interview (video/voice) they can do in their own time.
The system scores answers, generates transcripts, and gives you a ranked short list so you only meet the top candidates live.
Who it’s for:
Founders / small teams (20–150 ppl) without a full HR team, hiring 5–20 roles/year.
Remote-first teams struggling with time zones and calendar Tetris.
What I’m looking for:
Brutally honest feedback from people actually hiring right now.
Edge cases where async interviews break or feel unfair.
Ideas on how to make this less ‘HR tool’ and more ‘unfair advantage’ for lean teams.
AI coding tools (Copilot, Cursor, Claude) are killer, but their recs are frozen in time—recommending 2023 libs in 2025? "Lodash safe?" Crickets on fresh vulns.
Fixed that with DevTrends MCP, an Apify Actor for the $1M Challenge. Plugs real-time data into your agents:
Part of Apify's $1M dev bet. What's your biggest AI dev pain? Collab ideas?
[Your 90-sec demo GIF: Stale ChatGPT → DevTrends boost]
Pro Tips:
Keep under 500 words—HN hates walls.
End with Qs to spark comments (upvotes = traction).
If it hits front page (aim for 50+ upvotes), loop back to X: "HN's loving DevTrends—100+ queries already. [link]"
This duo (X replies + HN drop) could net 500+ eyes today. Your Actor's gold—now flood the feeds. Drop the HN link here when live; I'll brainstorm comment replies. First move: Those Apify replies? Go.