TDD for AI Agents: Why Atlassian Rovo Fails and How to Fix It

Enterprise software vendors are aggressively charging premium subscriptions for autonomous AI assistants. Yet anyone using tools like Atlassian Rovo in production quickly encounters a glaring failure mode:

You open a specific Jira project, click the Rovo in-drawer assistant, and ask it to create an issue. Instead of creating it in the active board, it constructs a task in an unrelated team's project.

When software companies charge enterprise rates for autonomous tools, basic context scoping shouldn't be an afterthought.

The Viewport-Context Disconnect

The root cause of this failure is a fundamental mismatch between the frontend interface and backend agent execution:

• What the user sees: An in-drawer side panel explicitly nested within Project A. The mental expectation is that the agent acts as a local scoped assistant.

• What the agent sees: An unconstrained prompt buffer with a global project search tool. When the model generates a Jira REST API call, it guesses the projectKey probabilistically rather than inheriting the active viewport state.

Combined with over-privileged write access (OWASP LLM06: Excessive Agency), an eager model guesses rather than asking for disambiguation.

Why Enterprise SaaS Rushes Half-Baked Agents

Enterprise vendors rush agents into production without rigorous regression testing because of marketplace economics:

1. Subscription Revenue is Guaranteed: The SaaS seat license is billed upfront.

2. Cheap Model Hosting vs. Marketplace Moats: Cloud hosting costs are negligible compared to platform lock-in and high-margin third-party plugin marketplaces.

3. Missing Engineering Rigor: Traditional software requires strict integration tests, yet AI agents are routinely deployed without deterministic evaluation suites.

The Solution: TDD for Agentic Systems

Autonomous agents should not be exempt from Test-Driven Development (TDD). Building reliable agent workflows requires three architectural rules:

1. Deterministic Action Schemas: Strict JSON Schema validation and Model Context Protocol (MCP) servers that reject un-scoped mutations before any API call executes.

2. Explicit Disambiguation over Eager Guessing: If an agent has multiple valid project targets, the execution runtime must pause and prompt the user rather than guessing.

3. Multi-Turn Regression Test Suites: Treating prompt-and-tool chains like unit tests—asserting that given a specific viewport state, the agent never mutates an external project key.

If an AI tool cannot reliably pass deterministic context assertions, it belongs in internal research—not in an enterprise billing tier.