AI Agents

Deterministic automation is the better answer when inputs, rules and outcomes are stable. A conventional chatbot may be enough when the job ends with a response. An agent becomes useful when the work contains judgement, unstructured material, changing context and a sequence of actions that cannot be described as one fixed rule set.

The distinction matters commercially. Calling every language-model feature an agent creates unnecessary cost and risk. We begin with the operating decision: what event starts the work, what state must be remembered, which result closes the run and which conditions must return control to a named human owner.

A production agent should not inherit every permission held by the person commissioning it. Read access, draft creation, approval and irreversible action are separate capabilities. Each tool receives the narrowest data scope and action scope that can complete the job; sensitive steps can require an explicit human decision or a second system check.

We document prohibited actions as carefully as allowed ones. The agent must know when a source is missing, when a request conflicts with policy, when identity or authority cannot be established and when an external system has returned an ambiguous result. A safe refusal or clean escalation is a successful outcome when the alternative is an unauthorised action.

A fluent answer can conceal a failed operation. Evaluation therefore starts with representative tasks, difficult edge cases, malicious or misleading inputs and known failure states. We check whether the agent chose the right tool, used the right source, respected the permission boundary, produced the expected record and stopped at the correct point.

The evaluation set is versioned with the workflow. Changes to models, instructions, tools, source material or business policy can alter behaviour, so a previous pass is not treated as permanent proof. Production monitoring watches completion, hand-offs, corrections, latency, cost and unintended actions separately; a single blended score would hide the failure that matters most.

The first release should solve a narrow, valuable job with a clear owner. It may begin in shadow mode, prepare drafts for approval or operate only on a controlled subset of records. This creates real evidence about quality, exception volume and integration cost before the agent receives broader access or handles a larger share of the process.

Production readiness includes more than a model and prompt. The written scope names the systems, data classes, tool permissions, model provider, expected volume, service limits, logging, recovery path, support window and acceptance tests. Pricing is supplied after those facts are known because the cost of a read-only research agent is not comparable with an agent authorised to change customer or financial records.