A video of an AI agent booking a meeting, reconciling an invoice, or clearing a support queue is easy to find. Getting that same agent to do those things on a Tuesday, against your actual systems, with inputs nobody scripted, is a different project. The gap between the two is where most of the budget goes.
Who this is actually for
Agent work pays off when the job involves judgement that shifts from case to case: reading a messy request, deciding which system to touch, and choosing between plausible next moves. Back-office operations, research triage, and support resolution fit that shape.
Plenty of tasks do not. If the work follows fixed steps, a plain script is faster, cheaper and far easier to reason about when it breaks. Adding a reasoning loop to a job a function already handles buys you cost, latency and a new class of failure. Knowing which side of that line you sit on is the first decision, and it is worth making honestly before any code exists.
What usually goes wrong
The recurring failure is scope. An agent handed broad permissions and a vague goal will act confidently and wrongly, and the trail of why it did so is often missing entirely. Without scoped tools, approval checkpoints where mistakes cost money, and traces of each decision, you have a machine that generates expensive surprises nobody can reconstruct afterwards.
Then there is reliability. Demos assume tools respond, inputs are clean and the happy path holds. Production assumes the opposite. Retries, fallbacks, cost ceilings and latency limits are not polish layered on at the end; they are the work.
What a trustworthy agent feels like
Done well, the agent has a narrow, named job and a defined boundary around what it may and may not do. It retrieves from your real data rather than guessing, and it remembers what it has already attempted. Risky actions pause for a human.
Behind it sits a test suite covering edge cases as well as the easy path, and a trace of every run that a non-engineer can follow. Costs are bounded, so a loop cannot quietly spend a week's budget overnight. And the code belongs to you, not to whoever built it, which matters the moment you want to change a tool or a rule.
The bar to hold a build against
Ask what happens when a tool times out, who approves the costly step, and how you would know the agent made a bad call three days later. If those answers are vague, the demo is doing the persuading rather than the design.
The full guide at Foundry 5 on production AI agent development in London covers the specific package every build ships with: how tools get scoped, where approvals sit, how evals and fallbacks are structured, and what observability and cost controls look like in practice.