Agent teams fail quietly, in five places at once.

A senior engineer wrote a piece this summer that every team running AI agents should read. He loved agent teams, multiple Claude instances splitting a feature and coordinating in parallel. Then he turned them off. Two reasons, and both are governance problems wearing an engineering costume.

The cost you can't see

A multi-agent run burns roughly fifteen times the tokens of a single chat, and the waste is not where you would guess. It is in scoping. Give a vague task to a big repo and the system spawns a swarm of sub-agents that each re-scan the same files, independently, before anything useful is written. You pay for four agents to rediscover the same code, and it is invisible until you go read what the agents actually did.

The numbers are not hypothetical. Uber burned its entire 2026 AI budget in four months, mostly on Claude Code, with power users spending $500 to $2,000 a month and a single two-hour session costing $1,200. Microsoft told an entire division to stop using Claude Code after a token bill ate the division's annual AI budget. One financial-services team ran 23 unattended sub-agents for three days and got a $47,000 invoice.

The failure that made him quit

The second reason has a name: the cascade error. An early agent makes a small, plausible mistake. The next agent builds on it. The next builds on that. Nothing fails loudly. By the time anyone reads the pull requests, the branches are stacked on a flaw, and untangling it costs more than the feature would have. One cited incident ended in data loss.

His own conclusion is the point. You do not have a parallelism problem. You have a scoping problem. Parallel agents are very good at doing a lot at once, which is exactly what turns a vague task into five times the wrong direction.

Where Signet fits, and where it doesn't

That is a governance problem, and it is the one we built for. Signet is a deterministic gate on every agent action, with a per-action budget and a tamper-evident record. Runaway spend becomes a metered, capped number you see in real time instead of discovering on the invoice, because an action over budget is blocked before it runs. Scope stops being a discipline you hope each agent remembers and becomes a sealed contract the gate enforces, so agents can only touch what their mandate allows and stop clobbering each other's work. And the destructive ending, the data loss, is the exact thing a deterministic gate blocks at the boundary, with a human checkpoint forced on the high-stakes actions rather than hoped for.

We will be honest about what we do not do. Signet does not read a pull request and tell you the logic is subtly wrong. That is code review, and you still own it. What we govern are the two things that made that engineer quit: the runaway spend, and the blind compounding that loses data while you are not looking.

If you have agent teams running and you cannot see the spend or prove what each one touched, that is the gap. We should probably talk.

Source: a June 2026 practitioner account, plus the publicly reported figures from the companies named.