“Agent” is the most overloaded word in the sector right now. To some it is a loop that calls tools; to others, an autonomous digital employee. That ambiguity is why so many projects start with impossible expectations.
Numbers before definitions.
Where we actually are
The 2026 AI Index from Stanford measures agents on OSWorld, a benchmark where the system operates a real computer. Task success went from 12% to 66% in a year. That is an enormous jump and, at the same time, a figure to read carefully: one in three attempts still fails.
On adoption, McKinsey’s State of AI 2026 survey — 1,719 responses, fielded between May and June 2026 — finds 62% of organisations at least experimenting with agents, but only 23% scaling them in any function.
That 39-point gap is not a technology shortfall. It is what happens when 66% reliability meets a business process that needs 99%.
The three places they break
1. The agent doesn’t know it failed. The dangerous failure mode is not the visible error, it is the step that looks like it went fine. An agent that misreads a field and proceeds confidently does more damage than one that stops.
2. Scope creeps by accident. OWASP moved excessive agency from sixth place in 2025 to third in its 2026 Top 10 for exactly this reason: we give agents persistent memory, credentials and tool access, then discover what they can reach.
3. Nobody knows what happened. Without per-step traces, debugging an agent is archaeology. And if you cannot debug it, you cannot improve it.
What is working
The deployments we see reach production share a pattern, and it is not the one the demo sells:
- Narrow scope, few tools. Five well-defined tools outperform thirty generic ones. The decision space shrinks and reliability climbs.
- A human at the point of commitment. The agent prepares, proposes and documents; a person approves the irreversible action. Sending the email, issuing the credit note, amending the order.
- Idempotency and retries. If a step can run twice without breaking anything, a 34% failure rate becomes a cost problem rather than a correctness one.
- A budget per task. Caps on steps, tokens and wall-clock time. An agent without a ceiling is an invoice without a ceiling.
- Continuous evaluation, not spot checks. A set of real cases that runs on every prompt, model or tool change.
The question that saves six months
Before building anything: what happens when it gets it wrong?
If the answer is “we catch it in the weekly review”, don’t deploy. If it is “the system detects it, rolls back and flags it”, you have a design. Most failed agent projects didn’t fail building the happy path — they never built the other one.
Where to start
The best first agent is not the most impressive one, it is the most boring and most repeated. Reconciling two systems that don’t talk to each other. Drafting the report someone rebuilds every Monday. Classifying and routing whatever lands in a shared inbox.
Tasks where the cost of error is low, volume is high and a verifiable correct answer exists. There, 66% is not a problem: it is 66% of work nobody does any more, with a 34% that gets reviewed.
Scaling from there is an engineering problem. Starting with the spectacular case is a budget problem.