Article

Migrate AgentCore file hooks before December 12

For teams running AgentCore file-policy hooks, identify affected harnesses, update event handling and preserve checks before the December cutoff.

Editorial illustration for Migrate AgentCore file hooks before December 12: an arrow connects an old service to its replacement. Not documentary evidence.

Teams with Amazon Bedrock AgentCore harnesses created before November 12, 2026, at 00:00 UTC must update hooks that still depend on file_operations before December 12, 2026, at 00:00 UTC. AWS says read, write and edit events replace that legacy input; newer harnesses already use them exclusively.

A harness is the managed loop that lets an AI agent choose and run tools; hooks attach checks or notifications at defined checkpoints. AWS’s November 19 walkthrough explains the broader Strands-backed update, whose rollout began November 11. This guide concerns existing file-policy and audit integrations, not every AgentCore application.

Find the affected harnesses and consumers

Start with GetHarness in each account and region you operate. Save the configuration and record createdAt, hooks, allowedTools and harnessVersion. Use the creation timestamp—not the last update time—to identify older resources.

Inspect the targets configured under beforeToolCall and afterToolCall. Follow each target through its handler, subscription filters and downstream audit mappings. Record who owns each dependency before editing it; the resource configuration alone cannot tell you what an external consumer does.

For a separate customer-data decision, see choosing Skyflow controls for AgentCore tools and models. That guide addresses what recipients receive, not this file-event migration.

Update event handling without widening permissions

Update context.toolName matching and downstream filters. Inspect native toolInput fields too: the legacy before-tool payload is translated, so a name-only change does not establish compatibility. Check the fields your policy actually uses; the notice does not supply a complete input-field migration map.

The permission list is separate. AWS still documents file_operations as an allowedTools compatibility selector granting all three tools: read views files, write creates or replaces them, and edit changes existing content. Changing that selector alone does not repair a handler.

Do not switch to ["*"] to solve this migration. The wildcard admits the current recommended tool set, which can grow. Keep explicit selectors when you need a fixed permission boundary.

Check the replacement events independently

During transition, legacy/new event pairs share toolUseId; either Lambda denial counts. Consequently, a working legacy check can conceal a missing replacement. Test new-only inputs, not just pairs.

For audit reporting, our recommendation is to correlate pairs within each harness, session, hook and phase. Retain raw records and decisions; never suppress policy evaluations or assume globally unique IDs. Do not depend on arrival order.

Keep the checkpoint unchanged when its purpose is prevention. Under AWS’s documented Lambda behavior, a before_tool_call denial skips the call while the loop continues. An after_tool_call denial retains the completed result and stops the invocation; it does not undo a write.

Use representative non-production files to check allowed and rejected actions for each permitted tool. Exercise target errors and timeouts against your intended failureMode; AWS says its default is deny. SNS and EventBridge are notification-only targets, so verify downstream receipt and processing rather than expecting a blocking decision.

Preserve all hooks when changing configuration

Separate a consumer-code change from a harness-configuration change. If the same target can process the replacement events, changing its implementation or filters may be sufficient. If you need different hook targets, prepare the complete intended configuration first.

UpdateHarness replaces the entire hooks list when you supply it. It does not merge one edited hook into the old list:

  • Omit hooks to retain the existing list.

  • Supply the full intended list to change it, preserving unrelated checks and notifications.

  • Do not send an empty list unless you intend to remove every hook.

  • If changing allowedTools, preserve every intended selector too; that list is also replaced in full.

Compare the saved and proposed configurations before submitting the update, then retrieve the resulting version to confirm the intended hooks remain.

Verify which version production uses

A successful configuration update is not the whole rollout. AWS creates an immutable version for each harness update. DEFAULT automatically follows the newest version; named endpoints remain on their selected version until explicitly moved.

Identify the endpoint your application actually invokes. If it uses DEFAULT, a harness update affects that route automatically—do not treat it as isolated staging. For a named production endpoint, review the new configuration, move the endpoint deliberately and confirm its target version.

Keep the prior configuration for comparison and recovery planning. Do not assume version pinning extends a service-wide event-removal deadline: configuration versioning and event compatibility are different contracts.

Budget the services behind the harness

AWS lists no separate charge for the harness itself, but model inference, Runtime, Memory, optional tools, observability, storage and network usage can still cost money. Lambda request and execution charges also matter when functions process hooks.

Do not translate fewer compatibility records into a promised bill reduction. Use AWS Cost Explorer or Cost and Usage Reports for billed usage; traces explain behavior, not the invoice. For the wider operating budget, see Amazon Bedrock coding-agent cost planning.