Inspiration
Agent builders here told me the same thing: their agents work in demos, but they're afraid of what they'll do in production. Online, teams describe agents that "fail quietly": the chat looks fine, but the action was wrong.
What it does
Senik is a simulated world for AI agents that take real actions. You connect your agent by its endpoint, with no code access needed. Senik reads its tools and rules, runs realistic and adversarial customer scenarios, and checks what the agent actually did in a sandbox, not what it said. It suggests a fix, reruns the same tests to prove it holds, and saves every failure.
How I built it
Solo, with Next.js and TypeScript, Anthropic and OpenAI models, an MCP recording proxy, Supabase and Vercel.
Challenges and what I learned
Routing a real agent's tool calls through Senik across serverless instances was the hardest part. The biggest lesson: an agent's words and its actions often disagree.
Built With
- ai
- nextjs
- react
- supabase
- typescript
