selfradiance
selfradiance on Hacker News
Interesting approach. The deterministic verification vs. LLM-as-judge choice is the right call — probabilistic safety checks on safety-critical actions is a category error.
One thing I've been thinking about: policy-based pre-execution authorization handles the prevention side well, but there's a complementary problem — what happens when an agent operates across trust boundaries where you can't predefine every allowed action? I've been exploring an economic accountability model (bond-and-slash, inspired by crypto staking) where agents post collateral that gets slashed on verified misbehavior. Prevention + accountability as two layers rather than one.
Repo if anyone's curious about the other side of this:
https://github.com/selfradiance/agentgate
Interesting project. Clean architecture (Next.js + SQLite + Drizzle, monorepo with Turborepo), and the SKILL.md format is becoming a real thing across Claude Code, Codex, OpenClaw, and OpenCode. The value prop is clear for teams managing lots of agent skills — structured editor, multi-model testing via OpenRouter, version history, one-click deploy. For solo builders who hand-maintain a few skills, it's probably overkill right now, but worth watching as the agent skill ecosystem matures.
Posted by
selfradiance on Thursday March 12, 2026 @04:07PM
from the show-hn dept.
https://github.com/selfradiance/restarules
RestaRules is an open standard that lets restaurants publish machine-readable conduct rules for AI agents. Any agent — voice, booking, concierge — checks a /.well-known/agent-venue-rules.json file before acting.
The rules cover things like: must the AI disclose it's not human? Which channels are allowed (phone, web, SMS)? What's the party size limit before a human gets involved? Are deposits...
The Zeigarnik effect point is underrated here. I've caught myself lying in bed running prompts in my head after a long session with Claude Code. The fix that's worked for me is treating each session like a pomodoro — hard stop, write down where I left off, walk away. The loop is real.
The Agent Council approach is interesting — having multiple small models debate in parallel and a judge synthesize feels like a more principled version of what people do manually when they cross-check answers between Claude, GPT, and Gemini. Curious whether the GSM8K gains hold up on less structured tasks where there isn't a single correct answer (e.g. summarization or open-ended reasoning).
Nice idea — prompt injection is one of those problems that gets hand-waved until it's in production. The CI/CD gating (fail hard on Critical) is the right call. Most security tooling only warns; blocking the pipeline is what actually changes behavior.
The .well-known/ suggestion from neilv is worth taking seriously. It's an IANA-registered convention that already has broad adoption — security.txt, robots.txt equivalents, app associations, and a growing number of machine-readable discovery files use it. Putting the discovery document at /.well-known/satproto.json instead of /satellite/satproto.json would immediately signal to anyone familiar with web standards that this is a protocol endpoint, not a page on your site. It also avoids the path collision problem that the satproto_root.json fallback is trying to solve — .well-known is reserved specifically so it doesn't collide with your content.
Appreciate the skepticism — the project looks like it required coding because it did. I just wasn't the one writing it. Claude Code wrote every line. I handled architecture, sequencing, security decisions, and verification. The shift isn't that coding stopped mattering — it's that the person directing the work no longer needs to be the one typing it.
Posted by
selfradiance on Wednesday March 04, 2026 @12:54AM
from the show-hn dept.
Posted by
selfradiance on Friday February 27, 2026 @06:19PM
from the show-hn dept.
https://github.com/selfradiance/agentgate
AgentGate is a small backend microservice that enforces stake-gated execution of actions.The idea is simple: as AI agents reduce the marginal cost of sending bids, API calls, negotiations, etc., systems designed around human friction become vulnerable to synthetic pressure.AgentGate requires:Identity registration with an Ed25519 public keyBond lock before action executionCryptographically...