Proxy vs tool-call gate: where should agent guardrails actually sit?
quiet_flare · reddit · 2026-11-03
The author surveys this week's agent guardrail tools (traffic proxies, secret-blocking extensions, MCP crypto identity, GitHub Actions audit logs) and argues the enforcement layer matters more than features: proxies see wire traffic but miss irreversible decisions made in the harness. They built Mainsheet on the Claude Agent SDK with per-tool-call gates (arg patterns, deny rules, budgets, human approval), taint tracking that blocks irreversible tools after untrusted output, and signed hash-chain receipts — while naming its limits (coarse session-wide taint, undocumented key custody). Asks practitioners where they put enforcement.
More from coding & agent
- Agent edits Qwen3.5-0.8B weights directly, lifting MATH-500 from 7.9% to 48.1% — giffmana · 2026-11-03
- Railway hack targeted AI keys: 'An API key is a credit card for compute' — alvelda · 2026-11-03
- Claude writes neural net weights from scratch: 959 programs, 100% accuracy on 10 algorithm tasks — ChengleiSi · 2026-11-03
- Anthropic and UK AISI report AI agent trying to push malicious code into real open-source repo — LingmingZhang · 2026-11-03
- Open-source herdr desktop app: every UI part is a file your agent can edit — philipvollet · 2026-11-03
- TalkPix ships remote MCP server with server-side spend limits for agents — No-Influence-161 · 2026-11-03