Latest in AI
AgentCore Strands Harness: How to Migrate Tool Policies and Hooks
AWS is upgrading the managed harness in place, but teams with file-access rules have two separate things to check before December 12.
Amazon Bedrock AgentCore is moving existing managed harnesses onto Strands. AWS detailed the update on November 19, after starting the rollout on November 11. You do not need a new harness, but you should check what your configuration lets AWS choose and whether your tool policies still recognise the new filesystem events.
Check what your harness leaves to AWS
An explicit model, system prompt or truncation policy still wins over the managed default. If you left those settings open, AWS can change the recommended model, instructions or context behaviour over time. The updated harness can summarise older conversation turns, for example, which matters if a long-running task relies on a constraint the user gave at the start.
Check allowedTools separately. With the wildcard ["*"], the harness can use both the recommended built-ins and configured tools that resolve, according to the AgentCore tools guide. That recommended set can change. If your support agent only reads troubleshooting notes, list read rather than letting the available capabilities grow with the default. Review effective per-invocation overrides too if your application supplies them.
Migrate the tool selector and the hook together
The old file_operations selector still works as a compatibility alias, but it grants read, write and edit. That is broader than a read-only agent needs. Keeping that selector also does not mean your hook will keep receiving events under the old tool name.
The lifecycle hooks guide says AWS emits both the old and new filesystem event names for harnesses created before November 12 at 00:00 UTC that have tool-call hooks, until December 12 at 00:00 UTC. The paired events share a toolUseId. Harnesses created from the November cutoff receive the new names only. If a dashboard counts both names as separate actions during the transition, it can double-count one file operation.
Update the handler to recognise the split tool names, and review its input validation. The compatibility event carries the old input shape, so renaming a branch in the handler is only part of the migration. If you update the hook configuration itself, send the complete list you want to retain. The hooks field on UpdateHarness replaces the list rather than appending to it.
Test a decision, not just the final answer
For the support agent, start with a harmless sample file and a request the agent should be allowed to finish. Then check that the deployed tool list has no write or edit capability. If a different workflow really does need to write, test its policy in staging with an allowed destination and a denied one, and check both the hook decision and whether the expected change happened.
Use a Lambda before_tool_call hook when a decision has to run before an action. A denial skips that call and lets the agent continue. An after_tool_call decision comes after the result exists and cannot undo a write. SNS and EventBridge are useful for reporting activity, but neither can make the allow-or-deny decision.
A regression check that reads only the final answer can pass even when the agent used the wrong tools. Save the expected tool list and policy outcomes alongside your expected answers. Those expectations give you a way to keep the convenience of managed defaults without approving every future change they bring.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders explain what changed in their products and what developers should do next.