The latest update to the Amazon Bedrock AgentCore harness swaps the internal harness for the open‑source Strands harness and bundles a set of managed defaults, new built‑in tools, and lifecycle hooks. For teams that already use AgentCore, the switch is transparent – the service automatically adopts the new defaults without requiring new resources or integration work. The practical impact is a richer, more cost‑effective agent runtime that can be steered with deterministic code at key points in the execution loop.
Managed defaults and built‑in capabilities
When a developer omits an explicit model, the harness now selects a recommended default that evolves with new model releases. A task‑oriented system prompt is injected automatically, guiding the agent to gather context, invoke appropriate tools, and verify its work before reporting completion. For supported Anthropic models, the harness caches the system prompt, tool definitions, and conversation history, allowing subsequent calls to reuse cached tokens and reduce latency and cost.
Context handling has been tightened. The harness monitors the token window and, once a configurable threshold is reached, summarizes older turns while preserving recent context. This proactive management reportedly reduces token usage by roughly a third compared with other harnesses. Large tool results can be offloaded from the immediate model context; the agent receives a preview and can fetch the full payload on demand, limiting repeated token consumption.
New tools extend the agent’s functional envelope:
todo_write– maintains a structured task plan that survives across invocations within the same session.subagent– spawns a fresh worker for a self‑contained sub‑task, returning the result to the parent without creating a separate managed resource.web_fetch– pulls live content from a specified URL to enrich the agent’s context.- Granular filesystem tools (
read,write,edit) – enable fine‑grained permissioning of file access.
Lifecycle hooks for deterministic control
Beyond the default behavior, the harness now exposes lifecycle hooks that run deterministic code at defined checkpoints in the agent loop. Practitioners can insert validation, logging, or policy enforcement steps before tool execution, after tool results are returned, or when a session is summarized. By anchoring non‑deterministic LLM output to predictable code paths, teams gain a clearer audit trail and can enforce governance requirements without altering the underlying model prompts.
Operational and security considerations
From an operations perspective, the upgrade does not introduce new resource types; the managed compute, scaling, and high‑availability features remain unchanged. However, the added defaults and tools increase the surface area of the agent runtime. Teams should review the permissions granted to each built‑in tool, especially the filesystem and sub‑agent capabilities, to ensure they align with the principle of least privilege. The proactive context summarization and offloading mechanisms reduce token consumption, which can lower cost but also alter the timing of data availability; monitoring should verify that summarization does not discard information required for downstream compliance checks.
Security engineers should note that the harness continues to run each session in a kernel‑isolated microVM, preserving isolation guarantees. The new hooks provide a convenient insertion point for custom security checks, such as validating tool arguments against policy or injecting additional observability metadata. Because the harness still relies on customer‑provided Model Context Protocol (MCP) servers, Gateway, Browser, and Code Interpreter integrations, existing IAM and network configurations remain applicable.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Adopting the updated AgentCore harness gives you a more capable default agent with less manual wiring, while the lifecycle hooks let you embed deterministic governance logic directly into the execution flow. Evaluate the default tool set against your security and compliance requirements, and consider enabling the hooks to enforce policy before tool calls. Keep an eye on the summarization threshold and offloading behavior to ensure critical context is retained for audit purposes. Finally, monitor token usage and latency metrics to confirm that the promised cost savings materialize in your workloads.


