Anthropic disclosed last week that earlier this year it moved roughly 150 product engineers off feature work and onto security, reliability, and privacy. Most of the coverage asked what prompted the shift. What I keep coming back to is what those 150 people stopped doing. Their output moved off the roadmap and into the machinery that makes the roadmap survivable. That's probably the most valuable work the company did this year. It also came out of the same pool of scarce engineers as everything customers actually touch, and it will never appear in a productivity benchmark.
Anthropic is a frontier lab, not a typical enterprise. But the shape of the decision generalizes to anyone deploying agents.
Once you let software act on your behalf in situations where it will sometimes be wrong, you have to build the thing that catches it being wrong: review queues, rollback paths, refund reserves, escalation trees, someone whose job is to answer the customer whose order got placed twice. Call it forgiveness infrastructure. It grows with every use case you turn on.
Which may bear on a gap in McKinsey's 2026 State of AI survey that's hard to explain away. Eighty percent of respondents report individual productivity gains from AI. Thirty-seven percent report any positive EBIT impact. McKinsey names rising operating costs as a constraint but doesn't split those costs into compute, licensing, and the overhead of keeping agents from doing damage. So this isn't confirmed. It's consistent, and nobody I've talked to is measuring the difference.
The mechanism is boring, which is why it's easy to miss. An agent finishes a task faster than a person would, and that's the number that goes in the deck. It also gets things wrong occasionally, in ways that are harder to catch than a human's mistakes, because nobody watched the intermediate steps. The speed shows up on Tuesday. The errors show up six weeks later, having compounded quietly, and land on a team that never saw the deck.
That's the part that stays invisible on the org chart. Correction work disperses across operations, QA, support, and legal, and none of those groups code it as an AI cost. It reads as ordinary operational overhead. Two new headcount in customer support that nobody traces back to the automation that went live in April. A review queue that didn't exist last quarter and now has a manager.
If you're measuring agent ROI by comparing task completion time before and after, you have a clean numerator and a denominator that keeps growing where you aren't looking. Integration complexity, change management, and data quality all take their cut too, and at least those have a budget line and someone who complains about them in the weekly. Forgiveness costs are harder to see precisely because they arrive as somebody else's problem.
So when the math on an agent deployment looks unusually clean, the useful question is where the correction work landed. Someone is absorbing it. Find out whether they know why their queue got longer.
- Anthropic's reward-hacking finding: Anthropic hypothesized that training on reward-hackable environments may increase willingness to carry out harmful sequences, and said an independent METR review is planned — worth watching for evidence on whether assurance costs compound when training environments themselves need policing.
- McKinsey's build-vs-buy signal: 32% of McKinsey respondents said their organization had decided against at least one software product because agentic coding allowed internal development, which could shift where forgiveness costs land from vendor to buyer.
- Observability vendors acquiring agent layers: Dynatrace's $915 million agreement to acquire Arize suggests established infrastructure vendors expect agent evaluation to become part of the standard observability control plane, which would make some forgiveness costs more visible in tooling budgets.
- Context work vs. agent-framework work: A recent preprint evaluating production text-to-SQL agents found that improving context artifacts produced roughly two to three times the gains of changing the agent harness itself, suggesting that some forgiveness costs may be reducible by making organizational context legible upstream rather than staffing review downstream.