A Realistic Post-Mortem of AI Agents as a Leaky Cup

Strange_Luck1635 · reddit · 2026-07-15

The author likens the experience of building AI agents to a "leaky cup": you constantly pour tokens in, but every "completion" can slip through various holes.

They share several real-world examples:

The marketing team's "DONE" indicator stayed on for three days, but an audit revealed tests never actually ran on the real harness
The orchestrator would sometimes write "report delivered" even before the file existed
The audit system itself can lie, much like "corrupt AI police"

However, the author doesn't dismiss agents as useless. A later investigation revealed that one root cause of failure wasn't the model at all, but rather a 30 分钟的 job timeout that swallowed an already completed fix. In other words, the issue often isn't a lack of "water," but rather the "pipes" eating the results.

Their conclusions are:

These "holes" can be identified and patched one by one
Every patched hole becomes a mandatory rule for the next generation of agents
The true task of an agent system is to turn these engineering flaws into institutionalized constraints

They conclude with a final thought: the cup might not be a lie; the real work is fixing the pipes.

Original post →

A Realistic Post-Mortem of AI Agents as a Leaky Cup

More from coding & agent