Hey Product Hunt 👋 I'm Hrithik, the maker of Anchor. What inspired it I've been building AI agents in production, and the same plumbing came up every time: a safe place to run code, memory that survives long tasks, budgets so a runaway loop doesn't burn money, traces you can replay, and some way to know "done" actually means done. Every team rebuilds this, and it's what slows down shipping. So I packaged it into one open-source SDK. What it does ⚓ Pluggable sandboxes: Heyo Computer, Docker, E2B or local, all behind one contract and tested by the same conformance suite (Daytona, Northflank, Vercel, etc coming soon) ⚓ Durable runs: conversations, workspaces and the agent's notes survive crashes, context limits and budget limits ⚓ Evidence, not vibes: every tool result is recorded, and passing tests are tied to the exact code they ran against ⚓ Guardrails and replayable traces, each with an off switch How the approach evolved My first version didn’t trust the agent at all. Every phase was gated: no moving from “fix” to “verify” without a passing test. The agents became rigid and literal. Even “review your work” turned into running tests for no reason. So I flipped the philosophy: trust the reasoning, verify the claims. The agent works freely, while Anchor quietly records the evidence (every tool result, every test run tied to the exact code it ran on). You decide where that evidence has to back up a claim. The result is agents that think like engineers and still can’t fake “done.” It's early alpha (0.1.x) and fully open source: 📦 npm i anchor-ai-sdk 💻 github.com/hrithikkoduri/anchor-ai-sdk I'd love your feedback: what's the piece of agent plumbing you've rebuilt the most? And if it's useful, a ⭐ on GitHub means a lot.

Forum Threads

p/anchor-sdk

•

1d ago

What’s the hardest part of getting an AI agent to production?

For me it was never the model. It was everything around it: where the agent runs code safely, what happens when it crashes at step 40, stopping a runaway loop from burning money, and knowing whether tests pass is actually true.

That s why I built Anchor. I m curious what s been the hardest part for you. Sandboxing? Memory? Costs? Trust?

View all

Anchor SDK

Build the agent. Not the plumbing.

Build the agent. Not the plumbing.

Forum Threads

What’s the hardest part of getting an AI agent to production?

Forum Threads

What’s the hardest part of getting an AI agent to production?