About
Agents are crossing into production. Verification hasn’t kept up.
Every team shipping an AI agent hits the same wall. The demo works. The eval scores look fine. Then the agent calls a refund endpoint twice, apologizes, and nobody finds out until a customer does.
The bottleneck in the agent stack has moved. Models are good enough. Frameworks are good enough. What is missing is the reliability layer — the tooling that answers “what will this agent actually do on a bad Tuesday with a flaky tool” before it happens in production.
Zeian Labs is building that layer. Zeian executes realistic tasks against your agent and its MCP tools, scores the full trajectory — not just the answer — and turns behavior drift into a failing CI check.
What we believe
- The trajectory is the product.
- Two agents can produce the same answer — one through a clean tool plan, one through retries and a lucky guess. Only one is safe to ship. Evaluation that ignores the path is grading the wrong thing.
- Tests belong next to the agent.
- Suites are YAML in your repo, reviewed in PRs, gated in CI. If a behavior change isn't deliberate enough to update a baseline, it isn't ready to merge.
- Real tools or it doesn't count.
- Mocked tools produce mocked confidence. Zeian exercises agents against actual MCP servers — including their timeouts, error shapes, and edge cases.
Zeian Labs is a small, independent studio. Zeian is early stage, building in the open with a small group of design partners. This site describes what exists and what is still being built — no theater.