Hacker News

Show HN: I built an "Agency Score" to measure how much your phone controls you

Hacker News - Thu, 02/12/2026 - 10:43am

Hey HN,

I kept trying phone detox apps but they all felt punitive – like I was fighting myself. So I built something different: a diagnostic tool that measures phone control as a single number (0-100).

Your Agency Score combines screen time, pickups, notification patterns, and app usage into one metric. Take a 3-minute assessment, get your score + a personalized action plan for your specific setup.

Free assessment, premium for ongoing tracking. Would love feedback on the scoring algorithm or UX. What metrics would you want to see?

Btw, you can try it on Mac as well. Link: https://apps.apple.com/us/app/superagency/id6757364702

Comments URL: https://news.ycombinator.com/item?id=46990204

Points: 2

# Comments: 0

Categories: Hacker News

Karpathy's Micro LLM in JavaScript

Hacker News - Thu, 02/12/2026 - 10:41am

Article URL: https://github.com/beachdevs/gptjs

Comments URL: https://news.ycombinator.com/item?id=46990184

Points: 2

# Comments: 0

Categories: Hacker News

Show HN: Tako AI – Agent for Okta With Natural language (zero hallucination)

Hacker News - Thu, 02/12/2026 - 10:38am

Hi HN,

Every week I watched Okta admins burn hours answering ad-hoc questions from security teams: "Who has access to Salesforce?", "Find all contractors with GitHub access who haven't used MFA in 30 days." The answers always involved the same painful loop: dig through a slow web console, chain API calls, correlate CSVs, write throwaway Python scripts. Repeat next week.

I spent 12 months building Tako AI to fix this. You ask a question in plain English, it returns verified data.

GitHub: https://github.com/fctr-id/okta-ai-agent

THE ONE RULE: Zero hallucinations.

In identity and access management, a wrong answer is worse than no answer. If an AI tells your CISO a contractor doesn't have admin access when they actually do, that's a security incident. Tako never "predicts" an answer. It writes the code to find the answer, executes it, and returns the raw result. Ask the same question twice, you get the same data.

THE HARDEST PROBLEM: Scaling to 107+ API endpoints

Most AI agents break down past 10-20 tools. They hallucinate parameters, call wrong endpoints, invent fields that don't exist. We went through five architecture rewrites over 12 months.

Each iteration: new LLM drops (GPT-4, Claude 3.5), we rebuild the agent, hit context limits, watch it snowball into gibberish. The breakthrough wasn't bigger context windows — it was precise context engineering. Instead of cramming 107 endpoint definitions into a prompt, the agent dynamically discovers the right spec for the task at hand. It reads a custom JSON API documentation file for the specific endpoint it needs, constructs validated requests, executes them. No hardcoded tools per endpoint. We're adding full CRUD operations next.

HOW IT WORKS:

Multi-agent architecture based on ReAct (Reasoning + Acting). Each agent has a narrow job:

• Router: analyzes your question, decides local cache vs live API • SQL Agent: queries local SQLite cache for bulk data (10k users in milliseconds vs minutes via API) • API Agent: handles live Okta calls • Synthesis Agent: merges everything into final verified report

The API Agent has a self-healing loop that surprised us. When generated code fails — wrong parameter name, rate limit hit, API schema changed — it traps the stack trace, feeds the error back to the LLM with context, and rewrites the code. We've seen it recover from Okta API changes we didn't even know happened yet.

PRIVACY & SECURITY:

Runs 100% locally in Docker. You bring your own LLM keys (OpenAI, Anthropic, Gemini, or Ollama for fully offline). Your employee PII never leaves your machine.

READ-ONLY by design. All generated Python and API code runs in a sandboxed environment. Every execution is automatically verified against security patterns before running — code is logged and available for audit, but you don't manually approve each query.

WHAT'S NEXT:

We see this as a platform, not just an Okta tool. The pattern (local cache + live ReAct agent + self-healing code execution) generalizes to any SaaS API. Google Workspace, Slack, Workday — same architecture, different spec files. Working on write operations with human-in-the-loop approval next.

What would you want AI agents to actually do for you in 2026? Where do you see this tech going beyond chatbots?

—Dan

Comments URL: https://news.ycombinator.com/item?id=46990146

Points: 1

# Comments: 0

Categories: Hacker News

Show HN: Got VACE working in real-time – 30fps on a 5090

Hacker News - Thu, 02/12/2026 - 9:45am

I adapted VACE to work with real-time autoregressive video generation.

Here's what it can do right now in real time:

- Depth, pose, optical flow, scribble, edge maps — all the v2v control stuff - First frame animation / last frame lead-in / keyframe interpolation - Inpainting with static or dynamic masks - Stacking stuff together (e.g. depth + LoRA, inpainting + reference images) - Reference-to-video is in there too but honestly quality isn't great yet compared to batch

Getting ~20 fps for most control modes on a 5090 at 368x640 with the 1.3B models. Image-to-video hits ~28 fps. Works with 14b models as well, but doesnt fit on 5090 with VACE.

This is all part of [Daydream Scope](https://github.com/daydreamlive/scope), which is an open source tool for running real-time interactive video generation pipelines. The demo was created in scope, and is a combination of Longlive, VACE+Scribble, Custom LoRA.

There's also a very early WIP ComfyUI node pack wrapping scope: [ComfyUI-Daydream-Scope](https://github.com/daydreamlive/ComfyUI-Daydream-Scope)

Curious what people think.

Comments URL: https://news.ycombinator.com/item?id=46989470

Points: 1

# Comments: 0

Categories: Hacker News

Pages