Sandclaw – A Safety First Agent
Three-tier agent isolation: thinking agent can't touch the internet, only gatekeeper can execute.

Pause agents mid-deliberation and edit responses before commit for compliance.
Enterprise AI teams, regulated industry developers
LangGraph · AutoGen · CrewAI
NSED v0.5.1 ships the human-in-the-loop control plane we've been building toward: operators can pause agents mid-deliberation, review and edit buffered responses before they're committed downstream, patch live config without restart, and auto-flag underperformers by score divergence.
The operator UI ships as zero-dependency vanilla JS — no npm graph, nothing a compromised package registry can inject into a control plane that has authority over live agent responses.
For regulated deployments: EU AI Act Art. 14, MAS AIRM, and FINRA 2026 all require demonstrable human oversight. This is the architecture that satisfies it — and you can audit the implementation directly from the source repo.
Three-tier agent isolation: thinking agent can't touch the internet, only gatekeeper can execute.
10-per-day trust scarcity is clever, but unclear if reputation actually guides agent selection vs. vanity metrics.
Vouch decay prevents stale trust accumulation in a crowded agent reputation space.
Deterministic challenge pipelines that bots solve in under a second, humans can't.
Authenticity vectors for agent identity, but protocol adoption unclear and competitive landscape undefined.
Browser review layer for terminal agents when most tools are CLI-only.