Why AI Agents Pass All Tests Yet Silently Fail in Production

·1 views

AI agents that clear unit and end-to-end tests are still breaking in production, delivering wrong or fabricated outputs without throwing any errors or exceptions. Engineering post-mortems point to four core failure types: semantic drift, state collapse, tool hallucination, and loop exhaustion — none of which produce invalid JSON or HTTP errors detectable by standard test suites. In one case, an e-commerce recommendation agent saw accuracy fall from 94% to 61% after its embedding index fell out of sync with a daily-updated product catalog, compounded by an unversioned model update from its provider. A fintech team's API-integrated agent passed contract tests for six months before a silent backend change broke production behavior undetected. Experts recommend embedding distribution monitoring, model version pinning, canary evaluations, and semantic output validation to close the observability gap between 'it ran' and 'it produced the right result'.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How Small Companies Can Reclaim Unused AI Quota to Compete Around the Clock

A developer essay published on DEV Community argues that AI tools like Claude Code have fundamentally shifted competitive dynamics by enabling companies to run software development continuously, without the constraints of human labor. The author contends that well-funded firms and those with surplus GPU capacity can compound their first-mover advantage every night while smaller rivals stand still. However, the piece identifies an overlooked equalizer: most companies on monthly AI subscriptions consume only a fraction of their paid quota during daytime hours, leaving nights and weekends entirely unused. The author proposes that small companies can narrow the gap not by outspending larger rivals, but by systematically deploying their already-paid-for quota around the clock. The core argument is that the real disadvantage for smaller firms is not a lack of resources, but a failure to fully utilize the AI access they already fund each month.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Build Auditable Session Revocation Into Go Server-Rendered Login Flows

A secure server-rendered login system in Go should manage session creation, verification, refresh, and logout under a single server-side policy, with password recovery explicitly revoking old sessions before creating new ones. The recommended design uses an opaque cookie backed by a server-side session record, keeping recovery tokens strictly separate from login sessions to simplify expiry and revocation enforcement. During password recovery, only one request should be allowed to consume a recovery credential even if multiple tabs submit the form simultaneously, with all others receiving a generic invalid-or-expired response. A defensible audit trail should log events in order — recovery requested, credential verified, password changed, sessions revoked, and new session created — so that incomplete state transitions can be detected and alerted on. Monitoring should focus on the age and count of recovery flows that changed a password without completing session revocation, rather than relying on raw 401 response counts, which mix unrelated failure types into a single noisy signal.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Progressive Profiling Should Update Users, Not Recreate Them

A technical guide for healthtech platforms argues that progressive profiling must update a single, immutable user record rather than creating a new one when profile details change. The user ID should serve as the stable aggregate key, while attributes like email act as verified lookup identities rather than primary keys. Each accepted profile change should be recorded as a separately authorized, auditable state transition to maintain a clear account history. The design separates operational profile state from audit evidence, keeping current data bounded while allowing event history to grow into tiered storage. Sensitive data such as password material and reset secrets should be deliberately excluded from audit logs, with deletion schedules aligned to healthcare compliance and incident-response requirements.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Monitoring SaaS Fixed Its Own Blind Spot After 47 Minutes of Silent Downtime

A monitoring platform called PulseWatch suffered 47 minutes of undetected downtime when its own internal watchdog component failed silently due to a blocking I/O call inside an async Python coroutine. The stalled event loop prevented the meta-monitor from sending heartbeats, causing the external validator to falsely report normal status while the primary database collapsed. Engineers resolved the issue by implementing a three-layer self-validation system built entirely on standard libraries, avoiding new dependencies under an 8GB RAM constraint. A non-blocking event loop watchdog was introduced to detect stalls exceeding 500 milliseconds and trigger an OS-level signal, adding less than 0.1 percent CPU overhead. A secondary fix replaced an unbounded metrics dictionary with a bounded LRU cache capped at 10,000 entries to prevent out-of-memory crashes.

0 comments Read more at DEV Community