LAS VEGAS — Amazon Web Services (AWS) has introduced a powerful lineup of capabilities aimed at radically simplifying and accelerating the development of AI agents, marking one of the most significant announcements at re:Invent 2025. With the debut of Reinforcement Fine Tuning (RFT), AgentCore Policy, and serverless customisation, AWS is targeting major bottlenecks that prevent companies from moving agentic AI from test phases into full-scale production.
The newly launched tools automate training workloads that today require deep machine-learning expertise and heavy infrastructure investments. By reducing these barriers, AWS aims to help organisations scale AI agents that perform routine business tasks reliably, affordably, and with domain-specific intelligence.
AWS Targets the Agent Training Bottleneck
For years, enterprises have experimented with AI agents but struggled to deploy them at scale due to lengthy experimentation cycles, engineering overhead, and unpredictable model behaviour. AWS says its newest capabilities directly address these issues.
According to AWS benchmarks, Reinforcement Fine Tuning delivers 66% accuracy improvements compared to base models. Companies like Collinear AI report dramatic productivity gains, with model experimentation cycles shrinking from weeks to days using the upgraded SageMaker AI tooling.
Dr Swami Sivasubramanian, Vice President of Agentic AI at AWS, highlighted the inefficiency behind today’s agent deployments. “Most companies reach for the largest models, but most agent tasks—like document lookups or calendar checks—don’t need that level of intelligence,” he said. “This leads to unnecessary cost, slower responses, and wasted resources.”
Amazon Bedrock Reinforcement Fine-Tuning: Turning General Models Into Specialists
The headline launch, Reinforcement Fine Tuning (RFT) for Amazon Bedrock, completely abstracts away infrastructure complexity. Developers choose a base model, upload invocation logs or a custom dataset, define a reward function, and the service handles the rest—automatically orchestrating the training pipeline.
At launch, RFT supports Amazon Nova 2 Lite, with more models scheduled to join the lineup.
Industry partners are already reporting significant uplifts. Phil Mui, SVP of Software Engineering for Agentforce at Salesforce, said AWS tests show accuracy improvements of up to 73% over base models.
He added: “RFT will extend what we can achieve beyond supervised fine-tuning, helping us deliver more precise, customised AI solutions to our customers.”
Sivasubramanian emphasised that the future of fine-tuning lies in quality, not size.
“A carefully curated dataset of 10,000 agent interactions can outperform millions of generic examples,” he noted. “It’s like turning a family doctor into a cardiologist—hyper-specialised and extremely effective.”
AgentCore Policy and Evaluations: Guardrails and Governance for AI Agents
AWS also introduced Policy, a new rule-based control layer within Amazon Bedrock AgentCore. It lets enterprises define natural-language boundaries that govern what an agent can and cannot do, including which APIs, Lambda functions, or third-party systems it may access.
For example, a policy such as:
“Block all refunds above $1,000 unless reviewed by a human.”
ensures sensitive workflows remain protected.
Alongside Policy, AWS unveiled AgentCore Evaluations, a monitoring suite offering 13 pre-built evaluators across correctness, helpfulness, tool usage accuracy, safety, and success rates. The tool can analyse live interactions and trigger alerts when metrics deviate—for example, if satisfaction scores drop 10% within eight hours.
This automated evaluation framework gives organisations a reliable way to ensure agent performance stays consistent in real-world conditions.
AgentCore Memory Brings Episodic Learning to Enterprise Agents
One of the most forward-looking capabilities announced is AgentCore Memory, which introduces episodic memory into agentic systems. Each interaction is stored with context, reasoning steps, actions taken, and outcomes. A secondary agent then analyses these episodes to detect patterns and optimise behaviour over time.
Sivasubramanian compared the experience to great customer service: “Your favourite restaurant remembers your name, your favourite dish, and where you prefer to sit. Agentic AI needs that same combination of short-term and long-term memory to feel natural and personalised.”
This feature is already proving essential for complex enterprise environments. S&P Global Market Intelligence implemented episodic memory across its distributed agent platform, Astra.
Helene Astier, head of Technology for MI Enterprise Technology and Sustainability, said managing state across hundreds of specialised agents had become increasingly challenging.
“As our agent landscape expanded, keeping context consistent was difficult. A unified memory layer was exactly what we needed,” she said.
A New Chapter for Enterprise-Grade Agentic AI
With these launches, AWS is signalling a strong push toward making agentic AI more accessible, more affordable, and far easier to operationalise. By automating fine-tuning, strengthening governance, and enabling long-term learning, AWS is positioning its platform as a foundation for the next generation of enterprise AI agents.
As businesses accelerate their shift toward automation and intelligent workflows, AWS’s 2025 announcements may mark a turning point—transforming agentic AI from a promising experiment into a mainstream business capability.
Read more: Qlik Partners with AWS to Bring Trusted AI and Analytics to the New AWS European Sovereign Cloud
