#failure-analysis
Every post tagged "failure-analysis".
49 posts · showing 31–40
-
Was It Retrieval or Reasoning? Diagnosing a Wrong Answer
A day spent rewriting the prompt for a wrong policy answer, when the policy document was never retrieved. Three checks in the right order, and one log field that separates "ignored it" from "never saw it."
-
Injection That Crosses a Handoff
The fetcher had no write tools and the analyst never saw a web page. The separation held perfectly — and routed the payload around itself, through the one field that carries prose.
-
The Timeout That Looked Like an Empty Result
The health check timed out, the handler returned an empty list, and the agent reported all clear while two services were down. An empty result and a failed check must not share a shape.
-
The Second-System Effect in Agent Deployments
The first agent works because it was scoped well. The second aims at exactly the cases the first one escalated — which is the residual that resisted automation for structural reasons.
-
Why "Just Add an Agent" Fails as a Product Strategy
Usage goes curious, then low, then the feature sits there — and the team upgrades the model, which doesn't help, because capability was never the problem.
-
The Use Cases Everyone Tries and Nobody Ships
Six applications get attempted everywhere and stall at the same four points. Each has a narrower version that ships — and it's usually less exciting than the pitch that funded it.
-
When the Requirement Itself Is Wrong, Who Notices?
The behavior was intentional, decided two years ago, documented somewhere nobody looked. Every stage faithfully executed a wrong ticket — and every automated check confirmed it, by construction.
-
The Handoff That Breaks Every Autonomous SDLC Pipeline
Chain agents across the lifecycle and the first handoff fails hardest — because it's the only one whose input isn't a written artifact, and every downstream check verifies against its output rather than the intent.
-
Output Went Up, Review Capacity Didn't
PR volume doubled and review capacity didn't, because review is human and got no faster. The number to watch isn't merge throughput — it's how fast approvals are arriving.
-
Upgrading One Agent in a System of Six
A contained improvement to one classifier made the supervisor above it drop half its escalations. No error, no failing test — because each agent's own tests passed against its own version.