Amazon Bedrock introduces agents: AI models that can complete tasks

On 26 July 2023, AWS introduced agents for Amazon Bedrock in preview. The concept extended foundation models beyond generating an answer: they could interpret a request, divide it into steps, retrieve information and call actions prepared by the organisation to complete a task.

That distinction matters. An assistant responds; an agent attempts to move towards an objective using tools. The capability created opportunities in service, operations and software, while requiring stronger controls than an informational chat.

From conversation to action

A language model can explain how to check an order. To actually do it, the system must identify the customer, query the right application, interpret status and possibly execute an action. Every step depends on data, permissions and external APIs.

An enterprise agent therefore combines a model, instructions, authorised knowledge, tools, validations and logs. Model quality cannot repair an ambiguous API or an undefined business rule.

Least privilege and reversible actions

An agent should not receive broad access for convenience. Checking order status does not require permission to issue a refund; drafting a response does not require permission to send it automatically.

Early implementations should favour reversible actions or explicit confirmation. Financial, legal or customer-impacting steps may be prepared by the agent and approved by a person. Limits on amount, volume, frequency and scope should stop or escalate exceptional requests.

Suitable first cases

In ecommerce, an agent might gather the context of an order issue, consult stock or approved documentation, draft a response, classify a request, find inconsistencies across channels or create a task in another system.

Evaluation must examine more than the final wording. Did the agent understand the goal, choose the right tool, send the correct parameters, respect access rules, interpret the result and stop when information was missing? Tests should include normal, ambiguous and adversarial cases as well as unavailable APIs.

Observability and improvement

Businesses need traces of the request, steps, tools, outcomes, time, cost and human approval. Useful metrics include resolution rate, steps per task, intervention rate, errors, time saved and satisfaction.

Sitelicon combines development, integrations, cloud and AI to build controlled processes. We begin with the objective, design permissions and validations, and measure results before increasing autonomy. A useful agent is not the one that does the most; it is the one that completes a defined task reliably and traceably.

Editorial note: originally published in July 2023 and reviewed on 11 September 2026 to preserve the continuity of our editorial archive.