News
3 min read
Amazon Web Services has made Bedrock AgentCore generally available, a platform that enables developers to build autonomous AI agents capable of independently executing transactions. The system includes built-in mechanisms for payment orchestration and spending limits, allowing agents to operate within predefined boundaries. Almost simultaneously, NVIDIA has rolled out improvements to Nemotron 3.5 Lightning, a language model specifically optimised for use in agent workflows.
Both announcements address the same domain: the practical deployability of AI agents that do not merely retrieve information or generate text, but also take actions, control systems and initiate payments. For developers and businesses building in this space, each offers a different piece of the required toolbox.
Amazon Bedrock AgentCore targets what has long been one of the biggest barriers to deploying autonomous agents: allowing financial actions to be carried out securely. The platform provides payment orchestration through which agents can independently make purchases or initiate payments, but always within limits set in advance by the administrator. These built-in spending guardrails are designed to prevent an agent from inadvertently spending large sums or acting outside its mandate.
The platform is built for scale. Whereas earlier agent frameworks were often deployed for small-scale or highly supervised use cases, Amazon is explicitly targeting production environments where large numbers of agents operate simultaneously. General availability means this is no longer a preview, but a service backed by corresponding SLAs and support.
Concrete pricing and technical specifications for AgentCore were not fully public at the time of writing, but the service sits within the broader Bedrock ecosystem of AWS, which is already used by large enterprises worldwide. For companies already running on AWS, that significantly lowers the integration barrier.
NVIDIA's Nemotron 3.5 Lightning is a language model from the Nemotron family optimised for speed and efficiency, with agent applications as its primary target area. The recent improvements are aimed at enabling the model to perform better in scenarios where an agent works through multiple steps, calls tools or makes decisions based on external information.
Within the world of agent orchestration, model selection involves a trade-off between reasoning depth, latency and cost. Large frontier models generally perform better on complex tasks, but are slower and more expensive per call. Lighter, specialised models such as Nemotron 3.5 Lightning seek to strike a different balance: fast enough for multiple consecutive steps in an agent loop, with sufficient capacity for practical business tasks.
NVIDIA explicitly positions the model as part of a broader AI enterprise stack, where it works in conjunction with tools such as NIM (NVIDIA Inference Microservices) for straightforward deployment. Precise benchmark figures for the latest version had not yet been independently verified at the time of publication.
The concurrent moves by Amazon and NVIDIA illustrate a broader shift in the AI market. In recent years, the emphasis has been on improving the models themselves. Now, a significant share of attention is shifting to the surrounding infrastructure: how do you deploy a model safely, scalably and cost-efficiently as an autonomous agent?
For developers, this means they need to build less themselves in the areas of security, orchestration and transaction management. Platforms such as Bedrock AgentCore take over those layers, allowing teams to focus on the logic and domain of their agent. The flip side is a growing dependence on large cloud vendors for critical parts of the agent stack.
For investors and policymakers, it is relevant that autonomous agents are now reaching the stage where they can actually spend money on behalf of users or organisations. This places new demands on governance, liability and auditability, areas where regulation and standardisation are still evolving.
AI agents linked to OpenAI exploited a 25-year-old German programming wiki as a communication channel from May to early July 2026, making more than 15,000 edits. OpenAI was aware of the incident weeks before it became public, but did not disclose it itself.
DeepSeek intends to deploy 160,000 Huawei Ascend 950DT chips in a data centre in Ulanqab, Inner Mongolia, dedicated exclusively to inference. Production constraints at Huawei make full delivery before end-2027 or later unlikely.
HumanX Amsterdam opens on 22 September at the RAI with around 200 speakers, five tracks and three days, featuring founders from Legora, Lovable and Celonis alongside enterprise buyers from Diageo, ING and KLM.
