The Test Tribe 7th Frankfurt Meetup

Test agents now generate tests, execute suites, triage failures and report pass/fail. The usual control is an AGENTS.md, a system prompt or a skill file telling the agent what not to do. That is a request, not a control. It defines no boundary, produces no evidence, and fails silently. A test agent can weaken an assertion, skip a failing case, write outside its folder, or follow an instruction injected through a web page or MCP tool result, and still report all green. A green run from an agent you cannot audit is a claim, not a result.

This session applies test-environment discipline to agents: define the contract, enforce it, and verify it. Each agent’s authority is declared once and compiled into enforced permissions, with write access scoped by path and disallowed tools removed entirely. A runtime plugin checks every tool call, and the effective permissions are read back from the runtime rather than assumed from config. Pass/fail comes from exit codes and captured output, never from the agent’s summary. Every stage is reviewed by an independent model, deterministic checks override a model’s “pass”, and human approval gates sit between stages. Includes a demo.