The enterprise agent illusion: Behind the AWS Bedrock AgentCore launch

I attended the AWS Bedrock AgentCore workshop today in Hong Kong. Eighty people in the room, mostly product managers and C-level. When they asked who actually builds agents, about five people raised their hands.

Yet the lab immediately forced them into the terminal to run CloudFormation templates and configure VPC subnets.

This gap is the real problem with AI adoption. We are handing cloud terminal commands to managers who just want to solve a workflow.

The AWS presenter noted during the introduction that 'running an agent is not equal to running it in production.' But they forgot that running a production-ready cloud container is also not equal to actual team adoption.

The big news from last month is that OpenAI is no longer exclusive to Microsoft. Amazon put fifteen million dollars into their latest fundraiser, and now GPT-5.5 and Codex are live on Bedrock.

This matters for enterprise. You can run OpenAI's best models entirely inside your own AWS VPC. Your data never leaves your network, and it won't be used to train future models.

You get unified billing and pay-per-token pricing on your standard AWS bill, bypassing the need for separate commercial agreements.

Technicals aside, the new Bedrock AgentCore Managed Harness is a massive validation. It mirrors exactly how we think about workflows, but built into AWS infrastructure.

Instead of writing custom orchestration loops, you declare everything in a single JSON config. It separates the setup into four distinct blocks:

First, the agent, which is just the model ID and the system prompt defining the persona.

Second, memory. It uses a user preference strategy to asynchronously extract context from past sessions and automatically injects it back in on the next call under the same user ID.

Third, skills. Instead of cramming every instruction into the prompt, you mount markdown files from S3 and let the model load them only when needed.

Fourth, the gateway. This is your connector to external APIs. You define what is allowed using an OpenAPI schema and lock it down with Cedar Policy. The agent can only call GET requests to fetch data, making it impossible for the model to accidentally trigger a write command.

AWS shared a case study from an OpenAI project called Heuristics Kingdom. Three engineers built a complete game in thirty-seven days with zero manually written code, generating fifteen hundred pull requests.

Their main lesson was simple: too many instructions means no instructions. When you flood an agent with text, it gets confused.

They succeeded by setting up clear environmental constraints: directory maps, tables of contents, and strict structural boundaries.

The engineer's job has changed. It is no longer about writing code. It is about designing the environment and the rules.

AgentCore is impressive, but it is built for the IT department, not the team operator.

An agent that requires VPC subnets and IAM configurations is too heavy for an SMB owner trying to automate their daily intake.

If the delivery of an AI tool requires more engineering hours to maintain than the manual process it replaces, the team will stop using it in three months.

Our approach at dConcept remains workflow-first. Map the manual bottlenecks, define the triggers, and establish the human review steps. Start small and only build out heavy cloud infrastructure when the scale justifies it.