Solo Developer Builds Kavdren, a Verification Layer for AI Agent Responses
Few-Ad-1358 · reddit · 2026-11-22
The author is building Kavdren, an independent verification layer for a pain point that grows as agents enter real business workflows: an agent can give a convincing answer that contradicts company policy or relies on stale information.
His example: a customer asks about refund eligibility, the AI confidently approves, the current policy says otherwise, and nobody notices until the customer complains. Rather than tweaking the prompt, retrieval pipeline or model, he proposes independently verifying the response before it goes through — checking it against approved policies and source evidence, flagging conflicting or stale evidence, and keeping an audit trail of the verification decision. He stresses this isn't claiming universal truth, only whether the answer is supported by business-authorized evidence and policies.
He asks engineers running agents in production how they currently detect policy violations, outdated retrieval context or unsupported claims — LLM-as-judge, deterministic validation, or human review.
More from coding & agent
- LangChain's Chase: don't let subagents edit the wiki, use a dedicated curator agent — hwchase17 · 2026-11-22
- Coders build with language, but the masses want visual building tools for each other — mrjonfinger · 2026-11-22
- Claude Code orchestrates Codex subagents to maximize token limits — majidmanzarpour · 2026-11-22
- Undergrads Build Safety-Gated Self-Healing Agent for Dockerized Apps — ExpressionOne8668 · 2026-11-22
- Powerful Tools Aren't Enough: Why AI Agents Still Fail at Tasks — Few-Psychology-933 · 2026-11-22
- gamepad-mcp: open-source MCP server drives games like GTA6 via real HID events — Dave_Maynor · 2026-11-22