Hi, author here! This essay came out of a much larger review I've been working on on coding-agent reliability, and a ton of time operating these systems and watching them break (a lot).
I've seen agent + fleet failure show up in lost or ambiguous state, duplicated retries, stale workers, weak ownership boundaries, non-authoritative observations being treated as truth, bad admission control, etc etc. In other words, many of the failures popping up look like distributed systems engineering problems.
I'd be interested to hear from others if they've found this to be true in their own systems, or any counterexamples including places where distributed-systems framing may be a poor analogy for handling agentic failure modes.
Hi, author here! This essay came out of a much larger review I've been working on on coding-agent reliability, and a ton of time operating these systems and watching them break (a lot).
I've seen agent + fleet failure show up in lost or ambiguous state, duplicated retries, stale workers, weak ownership boundaries, non-authoritative observations being treated as truth, bad admission control, etc etc. In other words, many of the failures popping up look like distributed systems engineering problems.
The longer technical review arXiv document is here: https://arxiv.org/abs/2608.13867
and I put in more practical/runnable material for auditing your own factory here: https://github.com/sjarmak/software-factory-reliability (mine still isn't perfect and I'm working through the audit now...)
I'd be interested to hear from others if they've found this to be true in their own systems, or any counterexamples including places where distributed-systems framing may be a poor analogy for handling agentic failure modes.