AWS previews Amazon Bedrock AgentCore
AWS previewed Amazon Bedrock AgentCore, modular services for runtime sessions, memory, gateways, browser use, code interpretation, identity, and observability.
By Hypler editorial team · Source: AWS.
Original source date: .
Hypler publication: . Hypler updated: .
Estimated reading time: 5 minutes.
Topics: AWS, Bedrock, agents, observability

The preview named separable agent components
AWS announced Amazon Bedrock AgentCore as a preview on July 16, 2025. The release named runtime, memory, gateway, browser, code interpreter, identity, and observability capabilities, and said they could work with models and frameworks inside or outside Bedrock. AWS documentation described a managed runtime, MCP-oriented gateway, agent identity functions, isolated code execution, browser capability, and OpenTelemetry-compatible observability. The 2025 preview does not settle current availability, regional coverage, or suitability for a particular workload. Its continuing value is architectural: it gives teams components they can examine separately.
Infrastructure does not choose the policy
An isolated runtime does not select safe tools. A memory service does not decide what may be retained. A gateway does not determine whether an API action needs human approval. Map each platform component to a policy covering data categories, caller identity, allowed tools, permissible effects, evidence retention, and an owner who can disable the integration. This avoids treating a managed service as a substitute for a control that belongs in the application. A browser session can be technically isolated while its agent submits a consequential form; that action still needs the authority, validation, and review expected of a human action.
Memory needs a lifespan and a named audience
Model access, task-session state, long-term memory, and tool credentials should remain distinct layers. Define what enters each one, who can read it, how it expires, and how deletion is checked. Memory becomes a hidden failure source when stale facts, another user's context, or old instructions appear in a later task. Agent identity helps make a caller visible, yet it does not allow a task to borrow permissions or context. A durable design makes those rejections observable, rather than depending on a model's ability to remember a boundary across changing prompts and tools.
A research assistant can stop before external action
Consider a hypothetical assistant that retrieves from an approved collection, extracts cited passages, and prepares a draft briefing. Its browser is limited to a fixed list of public domains, it has no credentials for systems of record, and its memory lasts only for the task while storing source identifiers rather than unrestricted transcripts. The output includes citations and uncertainty labels but cannot send mail, create tickets, or publish a page. An inaccessible or untrusted source produces a reported gap. The workflow does not seek a workaround, reuse context, or broaden its access.
Failure tests show whether the pieces join safely
Test session expiry, memory deletion, source filtering, unauthorized tool calls, malformed parameters, policy changes during a task, and browser pages containing adversarial instructions. Include partial execution and a retry after network failure. The expected result shows whether an action occurred, prevents duplicates, and preserves enough evidence for a reviewer to choose the next step. Before production records, use synthetic fixtures for retrieval, isolation, authorization, and observability. Reject uncited answers, calls outside the allowlist, missing audit events, and state changes made without an approval token.
A preview calls for a restrained rollout
A limited release should use a small audience, read-only or reversible effects, rollback, and a named incident owner. Preserve configuration versions and test results so later upgrades can be compared with the previous setup, and check current AWS documentation before relying on availability, regions, limits, or upgrade behavior. Hypler's analysis is that modular infrastructure provides useful architecture vocabulary, not an authorization system. Each expansion of access should follow evidence from that workflow's acceptance checks, while prior configuration and results remain available for comparison.
Hypler analysis
Hypler’s view
Modular infrastructure can expose agent boundaries that products often blur. It still cannot decide what a workflow may retain, which tools it may use, or what business effect a person has actually authorized.
Hypler technical analysis
Engineering notes
Isolated runtime, memory, gateway, browser, code-interpreter, identity, and observability services separate concerns that agent products often blur. Their presence does not create a safe workflow by itself; policy must define how identities, sessions, tools, retained state, and external effects join for one request.
What to inspect
- Session isolation
- Inspect session creation, tenant or task binding, cleanup, cross-session access attempts, and whether browser or interpreter state survives beyond the intended task.
- Memory policy
- Record what is stored, retention and deletion behavior, read audience, provenance, and tests for stale or unauthorized context affecting a later task.
- Gateway trace
- Follow one tool call from model request through identity, policy decision, tool execution, observability event, and final receipt or error.
Scope and limits. This was announced as a preview in selected regions. It does not determine an application's acceptable data retention, tool permissions, or business-action controls. Platform components must be assessed in the data and authority context of each workflow.
Terms and definitions
- MCP
- Model Context Protocol, an open protocol for connecting AI applications to tools, data sources, and prompts through defined interfaces.
- Tool Calling
- A model interface pattern where the model requests a structured operation from an external tool.
- Least Privilege
- A security principle that grants only the minimum access needed for a specific task.
- Execution Evidence
- Records that show whether an approved action was attempted, what inputs and targets it used, and what result occurred.
- AWS previews the Amazon Nova Act SDK
AWS previewed the Amazon Nova Act SDK for browser automation agents, including IAM, S3 data controls, a browser tool, Python scripting, and testing support.
- Microsoft previews Deep Research in Azure AI Foundry
Microsoft announced a public preview of Deep Research in Azure AI Foundry Agent Service. The company described an API and SDK offering based on OpenAI's research capability, with Bing Search grounding and source-backed output.
- OpenAI introduces ChatGPT agent
OpenAI announced ChatGPT agent, a product feature that combined web interaction, research, code execution, and document work in a virtual computer environment.

Original source
Historical source briefing with Hypler engineering analysis. Source details reflect the original publication date; implementation examples are hypothetical. Technical analysis expanded October 24, 2026; this does not change the original source event or establish a new source verification. Editorial prose revised October 26, 2026; original event, publication and source-check dates are retained.