Agent demos are convincing because the happy path is short. Production is where the fourth step fails and nobody knows what the system did with the first three.
Narrow beats autonomous
The agents that survive contact with users do one job with a handful of tools. Triage a ticket and draft a reply. Reconcile an invoice against a purchase order. Pull three reports and produce a summary a human signs off.
General assistants that plan their own work are fun to watch and hard to operate. Every extra capability multiplies the ways a run can go sideways.
Tools are the interface that matters
Most agent failures are tool design failures. Vague names, overlapping responsibilities and error messages the model cannot act on. A tool should do one thing, describe exactly when to use it, and return a message that explains how to recover when it fails.
Plan for the step that breaks
Set a step budget and a wall-clock limit. Make every action idempotent so a retry does not send the email twice. Persist the state of a run so it can resume instead of restarting from the top.
When an agent gets stuck it should stop and say so with the work it completed, not improvise. Silent failure is the behaviour that erodes trust fastest.
Decide what it may do alone
Reading data, drafting and proposing are safe to automate early. Sending, paying, deleting and anything a customer sees should stay behind a confirmation until the accuracy numbers earn the promotion.
Watch the runs
Log every step with its inputs and outputs and give your team a way to replay a run. When someone reports that the agent did something strange, you want the trace, not a guess.
