Most agents are binary: budget available, full capability; budget exhausted, failure. The failure is rarely graceful — it's an error mid-workflow, a half-finished research task, a scheduled job that produced nothing and said less. Users experience it as flakiness, which is the reputation that kills internal agent adoption faster than any single outage.

Resilient systems degrade instead of failing, and the veridive toolset has a designed-in degradation path: a free tier of real capability — library search and listing, transcript pulls from imports, credit checks, pings. An agent that knows the path stays useful at zero budget; it just gets humbler.

The pattern is three standing instructions. get_credit_status as the sensor: the agent reads the budget state at decision points, not after failures. library_search (with library_list and free transcript pulls) as the fallback capability: when the budget is thin, answers come from the already-imported corpus — which for mature libraries answers a surprising share of questions. ping as the health layer: distinguishing "no budget" from "no connection," because the two need different escalations.

The agent's contract upgrades from "works until it doesn't" to explicit modes: full (deep tools), degraded (free tools + a note about what a deep run would add), and offline (notify). Users trust agents whose failure modes are designed; that trust is what adoption is made of.