The Dependency the Agent Invented: A Myth-Busting FAQ

A common misconception among teams is that a successful, error-free loop from an agentic agent confirms a shared understanding between the agent and the repository. However, this is a myth. A successful loop merely demonstrates that the code agrees with itself, not that external dependencies like Redis, object storage, or a message queue were properly accounted for.

Engineers often interpret a multi-step trace of file reads, patches, test commands, and summarized patches as evidence of a negotiated design. But this sequence more closely resembles a hiker familiarizing themselves with the scenery than a committee recording dissenting opinions. Green tests merely validate the code's internal consistency, not the existence of external services or resources.

Another fallacy is equating a successful tool call with the intended side effect. A zero exit code merely indicates the process completed, not that the external object described in the prompt was mutated. A formatter can succeed even on a file that should not exist, and tests can pass against a stub the agent has just created. This analogy is akin to a shipping label printing without verifying the crate contains the correct items listed on the invoice.

Furthermore, reading a file once and assuming the agent is grounded for the remainder of the loop is misleading. A transcript represents a snapshot of the codebase, which can quickly become outdated as later writes introduce new modules unseen by the initial read. Grounding, then, requires a series of checks after each batch of writes, comparing the current state with the contract, rather than relying on a single observation at the loop's start.

Lastly, having a large free-tier token budget does not eliminate the need for an assumption budget. Token volume only reflects how much text flowed through the loop, not how many new runtime names were authorized. It's crucial to allocate separate token allowances, ensuring that the final pass verifies both the contract file and the diff against the working tree, even if the initial summary appears complete.

The proposed solution is to implement a local checker that acts as a customs form, scrutinizing the agent's actions before the loop begins. This script reads a unified git diff, then rejects any service names, environment prefixes, ports, or imports not explicitly listed in the contract file. By committing this contract before the loop starts, teams can prevent undeclared dependencies from slipping through, ensuring a more controlled and auditable development process.