The quiet failure mode has arrived

Read this week's board as a single document and the argument is not that coding agents got more capable. It is that their failure mode moved. The old failure was loud: a crash, a red build, an error you could see. The new one is silent: a run that is green, a month of behaviour that was never what you asked for. A developer writing on dev.to described agents that ignored their instructions for four weeks while nothing errored. Another traced four different coding agents all pinning a plugin to a commit SHA while a branch with the same name quietly won the resolution. Both are the same bug class — the system did exactly what it resolved, and none of it was what you specified. That is a much harder problem than a crash, because there is no alarm to fire.

The fix being sold is governance, and it is worth reading sceptically. The big platforms have started shipping determinism layers, confidence layers and safety-and-governance layers as a series of posts. Treat those as a roadmap, not a guarantee. A section heading called 'determinism' is not determinism; you cannot buy the property that your agent will do what you meant by adding a layer someone else wrote. The only honest determinism is the one you test for, and the test is the part nobody is shipping.

There is a second, less obvious signal hiding in the infrastructure bucket, and it is about capacity, not capability. Postgres 19 was delayed because AI reviewers found more bugs than humans could fix. Read that carefully: review throughput is no longer the bottleneck, remediation is. Every team adding AI code review this quarter should treat that headline as a forecast of their own pipeline, not as a curiosity from someone else's project. When your reviewer gets good enough, the queue of found-but-unfixed becomes the thing that sets your ship date.

And then there is cost, which is the constraint most teams are actually engineering against. The vendor case studies say 76x cheaper at Asana, halved costs at LegalOn, half the research time at Parallel. All three are the same shape: cost fell because work moved to a cheaper tier or a narrower model, not because inference got cheaper on its own. The 76x number is scoped to browser tests with no published denominator, so treat it as a routing decision that happened to work. If your own bill is flat while everyone else's drops, the missing piece is almost always routing — not a better model, not a discount, but knowing which task can afford to get dumber.

Two open-source stories bookend the week and they should be read as procurement, not as migration. Munich is dropping proprietary software again — and, as the headline itself admits, not on the desktop. The Dutch tax authority cancelled a Microsoft 365 rollout and pointed at open source. Both are sovereignty and licensing decisions, deliberately scoped to servers and backends. They are evidence about vendor trust, not evidence that the hard part got easier. The desktop problem is still there, unsolved, which is why these announcements keep arriving with a qualifier attached.

What I am watching next week: whether anyone publishes the test that accompanies their determinism layer; whether the remediation queue, not the reviewer, becomes the talking point; and whether a real benchmark with a published denominator lands for the cost claims. The through-line is the same. The interesting questions about AI tooling are no longer 'can it do X'. They are 'what did it actually do, what will it cost, and who fixes it when it is wrong'. That is an operations question, and it is the whole job now.