Monitoring and rollback account for £45,000 of a £75,000 agent build

The reliability figures check out. 0.95 to the fifth power is 0.7738, and 0.95 to the tenth is 0.5987 ^[1]. Multiplying per-step rates that way assumes each step fails independently of the others. If a bad retrieval at step two feeds every later step, the assumption is wrong and the product is the wrong model of the chain. For the number to transfer to a real deployment, the 95% has to be measured on your own steps against realistic inputs, which a demo running the happy path never does ^[5].

Run the same product backwards and it becomes a design budget. To reach 95% end to end across ten sequential steps, each step has to be right 99.49% of the time ^[2]. The post prescribes architecture, not prompt edits: shorten the chains, verify between steps, and design each step so failure is detected rather than passed downstream ^[5]. A verifier inserted between steps is itself a step that can fail, so the count goes up before the success rate does.

Token spend follows from the same number. The post notes that spend scales with retries and chain length, so a failing agent costs more than a working one and a looping one costs a great deal more ^[9]. Retrying a failed ten-step run from the top takes 1.67 attempts per success on average, roughly 67% more tokens than a single clean pass ^[3].

On the build side, the £45,000 between the floor and the ceiling of the post's largest band is what it charges for monitoring and rollback done properly ^[4]. That increment alone is 3.75 times the ceiling of the simple internal agent band ^[5]. The post does not say how many projects the bands are drawn from.

Evaluation is the line most likely to be absent from a quote. Conventional software is tested against expected output, while an agent produces different output each run and can be correct in several forms, so the test artefact is a graded set of realistic cases with defined acceptance criteria, run automatically on every change ^[7]. The post's guidance is to expect that harness to take as long to build as the agent ^[8]. At the top band, that is about half the labour.

Most business processes have no rollback, and an agent that updates a customer record has changed something other systems have already read ^[6]. Email has no unsend, so any compensating action has to be written by hand. The post's observation is that building reversal is often more work than the action itself, and that it gets deferred and then discovered ^[6].

The deployments the post describes as working share one shape. The agent handles volume and a human handles judgement: triage with a draft response attached, document extraction a person confirms before it is committed, reconciliation that flags exceptions instead of resolving them ^[14]. Where the process has no stable structure, the post says, "An agent given a task with no stable structure will produce plausible output indefinitely without ever being right, which is worse than failing visibly" ^[12]. That is the case the eval harness exists to catch, because the output looks fine.

The compliance change sits on a fixed date. Article 50 transparency obligations under the EU AI Act took effect on 2 August 2026 as originally scheduled, and they require disclosure when a person is interacting with an AI system ^[10].