Amazon Bedrock Adds Agent-Led Migration Guidance
AWS published Amazon Bedrock migration guidance on November 12, giving coding agents a migration procedure that starts with discovery rather than production changes. The addition targets customer-managed knowledge bases, which supply application answers with retrieved documents. Its boundary matters: the agent helps prepare a replacement, while the team remains responsible for deciding when it can serve users.
The guide sits inside the amazon-bedrock skill, a package of task instructions and references for coding agents. It starts with read-only assessment and stops at a reviewable plan by default. Optional implementation creates a managed replacement alongside the original or produces infrastructure as code, configuration files deployed through an existing pipeline.
That creates a concrete choice for teams already running document-backed applications. They can keep managing retrieval infrastructure or evaluate AWS’s managed alternative without immediately abandoning the existing system. The central question is whether reduced infrastructure work justifies changed retrieval behavior, application updates, and a fresh acceptance process.
Amazon Bedrock migration begins with a reviewed plan
The announcement changes how teams prepare a migration, while leaving production acceptance in their hands.
According to AWS’s migration guidance, the agent examines the existing configuration before proposing changes. The workflow distinguishes settings that carry over, settings needing adjustment, and configurations without a supported managed equivalent. Blocked configurations are excluded rather than silently converted.
That classification gives reviewers something more useful than a generated deployment file. A replacement configuration describes what the destination contains. A mapping also explains what happened to the source system’s assumptions, including the ones that cannot survive unchanged.
For an engineering manager, those are different review tasks. Checking that a resource exists establishes provisioning success. Checking why a setting changed establishes whether the proposed destination still fits the application.
The distinction becomes especially relevant when the existing system has accumulated decisions across several owners. An infrastructure engineer might understand storage choices, while an application developer understands filters. Neither necessarily knows why an earlier team selected a particular document processing approach.
A written migration plan provides a place to bring those decisions together. Its value depends on whether reviewers can trace each proposed change to an existing requirement. An unexplained substitution deserves investigation even when the destination configuration passes a syntax check.
The parallel replacement also changes the conversation about risk. Reviewers can assess a candidate system while retaining a working reference. That makes disagreement easier to resolve through comparison instead of speculation about how a replacement will behave.
It does not eliminate the need to define success beforehand. A team that starts testing without acceptance criteria can keep discovering new reasons to approve or reject the destination. Agreement on representative questions, access expectations, and acceptable failures should precede the final decision.
The announcement describes a workflow, rather than publishing independent evidence that migrations consistently succeed. Readers should therefore distinguish a documented procedure from measured production outcomes. A clear process improves reviewability; its existence alone establishes neither retrieval quality nor business value.
For buyers, that makes the plan an early deliverable worth evaluating. Does it expose missing information? Does it identify decisions requiring an owner? Does it separate a technical incompatibility from a preference that someone can reconsider?
Those questions keep the discussion attached to the application. They also explain why an agent’s willingness to stop matters. Producing a complete-looking answer would be less useful than identifying a dependency that prevents a responsible migration decision.
Infrastructure teams face a different ownership decision
Managed retrieval moves operational responsibility toward AWS, but application owners still have to define useful behavior.
AWS’s managed knowledge bases documentation describes service-operated ingestion, storage, indexing, and retrieval infrastructure. Defaults include managed embedding and reranking models. Embeddings represent content numerically for matching, while reranking reorders candidate results for relevance.
The customer-managed route leaves teams choosing and operating the underlying stores. That distinction gives the migration story its practical stakes. Organizations are deciding which parts of their retrieval system justify direct control and which parts consume maintenance effort without serving a distinctive requirement.
For a platform team supporting several applications, reducing repeated infrastructure work has an obvious appeal. Every separate storage configuration introduces decisions about capacity, monitoring, permissions, and ownership. Standardizing those decisions can free attention for application behavior.
However, that argument depends on the workload. Some teams deliberately maintain direct access to their retrieval infrastructure because they need particular controls. A managed destination should be assessed against those requirements rather than treated as an automatic architectural upgrade.
The pressure falls on engineering organizations to explain their existing customization. A setting that once solved a specific problem might still be essential. It might also remain in place because nobody has revisited the original constraint.
Migration planning creates an opportunity to distinguish those cases. It turns an abstract preference for flexibility into a concrete review of what the application actually uses. That review can justify either moving or staying.
Consider an illustrative support assistant searching product documentation. Its users care whether the answer reflects the correct product version and policy. They rarely care which storage system located the passage.
For that application, operating the retrieval store is a means to an end. If the managed alternative meets those requirements, infrastructure ownership becomes easier to reconsider. If it loses a necessary filtering behavior, operational simplicity does not settle the decision.
Knowledge ownership also remains separate from infrastructure ownership. AWS operating a store cannot determine which internal policy is authoritative. It cannot resolve contradictory documents merely because both were successfully indexed.
This distinction applies across personal and team knowledge. Information quality and retrieval infrastructure address related problems, but replacing the latter does not repair the former.
Financial claims deserve similar discipline. Less infrastructure work can have value, yet savings depend on actual usage and the resources being replaced. A team should evaluate overlapping operation during migration alongside its expected steady-state workload.
The defensible buying question is therefore specific: which responsibilities disappear, which remain, and which become more important? Answering that question requires application owners and infrastructure owners in the same review. Neither group can establish the full case alone.
Bedrock managed retrieval changes the application contract
A destination that accepts requests can still behave differently from the system it replaces.
The most immediate example is the retrieval request itself. Managed queries use managedSearchConfiguration, while the customer-managed request shape uses vectorSearchConfiguration. These fields specify search behavior, so changing the destination identifier alone does not complete the application update.
AWS’s query configuration documentation also identifies differences in metadata filtering. The managed path does not support the startsWith and stringContains operators. Reserved system metadata uses underscore-prefixed names, which can affect application logic that reads those fields.
These details turn Bedrock managed retrieval into a compatibility question as well as an infrastructure decision. A field used only in an internal debugging script has different consequences from a filter enforcing application scope. Reviewers need to know where each dependency appears.
Imagine a support application that narrows results using a product-family prefix. If that behavior depends on an unsupported operator, a successful resource deployment leaves an unresolved application requirement. The team needs an explicit alternative and evidence that it preserves the intended boundary.
That example illustrates why generated configuration deserves review alongside caller code. Infrastructure files describe the destination, but application code expresses how users interact with it. Reviewing either in isolation leaves part of the migration unexamined.
Result counts require similar care. AWS documents numberOfResults as a maximum rather than a guarantee. An application that treats fewer results as a service failure could misinterpret a valid response.
The broader lesson is that an operation name does not define the entire contract. Inputs, defaults, metadata, ordering, and failure handling all influence behavior. Familiar naming can make a change look smaller than its consequences.
Microsoft provides a useful comparison without becoming the main opponent in this story. Its agentic retrieval documentation describes a multi-query pipeline that combines retrieved material for application use. Optional model-based query planning is marked as a preview capability.
That comparison shows another cloud provider organizing retrieval around a higher-level knowledge interface. It does not establish equivalent features, migration effort, or performance. Those conclusions would require a separate workload-based comparison.
For existing AWS customers, the immediate decision remains between their current retrieval arrangement and the proposed managed replacement. Switching providers would introduce additional questions and distract from evaluating the announced migration path.
Amazon Bedrock migration therefore deserves an application compatibility review before anyone treats it as routine resource creation. The useful output is a list of changed expectations, with each one assigned to code, configuration, testing, or an explicit decision.
That approach also improves accountability. When a result changes, the team can investigate a documented assumption instead of asking whether the agent generally did a good job. Specific expectations make both approval and rejection easier to defend.
Permission-aware search still depends on verified identity
Managed document filtering cannot replace the application’s responsibility to authenticate users.
An access control list, or ACL, records which identities may access a document. AWS’s ACL-aware retrieval documentation states that managed retrieval filters using identity supplied by the application. The service does not authenticate that end user.
That boundary has practical consequences for migration review. A search result can be relevant to a question and still be inappropriate for the person asking. Retrieval acceptance must therefore include questions about access, not just whether the expected document appears.
Consider an illustrative employee assistant covering general policies and restricted personnel material. A broad administrator account might retrieve both correctly. That result says little about whether an ordinary employee receives the appropriate subset.
Testing through a privileged development identity can hide this difference. The application’s actual identity flow needs its own review, including how it obtains and forwards the user identifier. A document permission model is useful only when the caller supplies trustworthy identity.
AWS documents email addresses as the user identifier for ACL matching. An address mismatch across connected systems can lead to missing results. The documentation also distinguishes connectors that support real-time permission checks from those that rely on supplied or synchronized information.
Those facts make identity consistency part of retrieval quality. A missing answer might reflect an access mismatch rather than weak search relevance. Treating every empty response as a ranking problem sends investigation in the wrong direction.
The analytical consequence is straightforward: teams need separate explanations for different failure categories. Missing content, denied access, stale permissions, and an irrelevant match should not collapse into a single retrieval failure label.
This also affects who must participate in acceptance. Search engineers can evaluate relevance, but identity owners understand account mapping and lifecycle rules. Content owners know which documents should remain restricted.
A migration that changes retrieval infrastructure can expose disagreements between those groups. That is useful information, even when it slows approval. An unresolved permission assumption already represents application risk, regardless of which system currently serves requests.
For coding-agent evaluation, the same principle applies to the evidence reviewers request. A convincing summary should identify the identity used for testing and the expected access boundary. A list of successful queries without that context leaves a material gap.
Bedrock managed retrieval can supply permission-aware filtering, but the team must evaluate the surrounding application path. That includes what happens when identity is missing, inconsistent, or associated with access that recently changed.
The strongest acceptance argument combines relevance and exclusion. Authorized users should find the information they need, while restricted material stays outside their results. Testing only the successful path establishes an incomplete picture of the replacement.
Retrieval quality needs evidence beyond a clean deployment
Provisioning success answers whether infrastructure was created; evaluation answers whether the application remains useful.
AWS’s retrieval evaluation metrics distinguish context relevance from context coverage. Relevance concerns whether retrieved text fits the question. Coverage concerns whether the retrieved material contains the information needed against an expected reference.
That distinction offers a better basis for migration discussion than a generic accuracy claim. A passage can concern the right topic without containing the detail needed to answer. Conversely, useful information can appear among results that also include distracting material.
For an illustrative policy assistant, finding the policy title is insufficient when the question concerns an exception. The retrieved material needs to include the exception itself. Otherwise, a later answer-generation step begins with incomplete evidence.
The same problem applies to technical support. A troubleshooting guide may be relevant to a device family but unsuitable for the requested version. Retrieval review should reflect the conditions under which users actually need an answer.
These are reasons to build acceptance around representative tasks rather than convenient demonstrations. Familiar questions make a system easy to showcase. Difficult questions reveal whether the replacement preserves behavior that users depend on.
Comparing the old and new systems remains useful, but agreement is not the final objective. The original can return an inadequate document too. A changed result deserves examination before anyone labels it a regression.
This creates an important role for expected evidence. Reviewers should know which source material supports the answer and why. That lets them distinguish a beneficial change from accidental similarity.
Evaluation should also preserve the difference between retrieval and answer generation. A model can write a fluent response from incomplete context. Reading only the final answer risks overlooking a retrieval weakness concealed by plausible language.
AWS separately lists answer measures including correctness, completeness, faithfulness, and citation quality. Those categories reinforce the need to inspect both the evidence supplied to a model and the response it produces.
The announcement does not provide independent workload benchmarks establishing a universal improvement from the migration guide. That limitation should shape the article’s conclusion. The available evidence supports a documented migration approach, while application-level benefits remain something teams must demonstrate.
Operational evidence belongs beside relevance evidence. A replacement might answer correctly during a quiet test yet behave differently under normal traffic. Latency, failure handling, and maintenance effort should therefore enter the acceptance discussion.
For an Amazon Bedrock migration, the most useful result is a review package connecting configuration decisions to observed application behavior. It should make disagreements visible rather than compressing everything into an overall pass.
That is where coding-agent assistance becomes assessable. The question is whether it helps produce understandable changes and reproducible evidence. A polished explanation without those artifacts gives reviewers little basis for approving production use.
Watch compatibility, workload results, and operating effort
The next evidence should come from supported configurations, representative evaluations, and sustained operation.
First, watch whether AWS expands or clarifies migration compatibility. The practical signal is documentation showing how previously difficult configurations map to the managed destination. General product announcements are less informative than specific answers about existing requirements.
This matters because compatibility determines the available migration population. A workflow that handles common configurations can already help some teams. Broader adoption requires clear treatment of the settings and dependencies that make other environments harder to move.
If those mappings become more complete, the case for evaluating migration strengthens. If important requirements remain unsupported, retaining the current architecture remains a reasonable outcome. A responsible plan should make that conclusion possible.
Second, watch for published results from representative workloads. Useful reporting would describe the source material, query selection, filtering requirements, and evaluation method. It would explain both improved outcomes and cases where the destination performed worse.
A headline accuracy figure without that context would leave too much unresolved. Buyers need to understand whether the tested application resembles their own. They also need enough detail to distinguish changed retrieval behavior from changed evaluation conditions.
Evidence showing useful answers and correct access handling would strengthen confidence in the managed route. Repeated problems around essential filters or identity integration would narrow its appeal for affected applications.
Third, watch operating effort after the migration period. Initial resource creation is visible and easy to demonstrate. The more consequential question is how much attention the application needs after ordinary content updates and user requests resume.
Teams should look for evidence about incident investigation, content freshness, permission changes, and ongoing ownership. Those experiences will help establish whether infrastructure consolidation reduces work overall or shifts it toward application integration.
These signals keep expectations proportional to the announcement. AWS has documented a path that helps coding agents prepare a move. It has not removed the need for engineering judgment about requirements, evidence, and production responsibility.
For readers considering an Amazon Bedrock migration, begin by identifying the behavior your current application must preserve. Ask for a read-only assessment that connects those requirements to the proposed destination. Bring application, content, and identity owners into the review before evaluating the generated changes. Then compare retrieval using questions grounded in real work, including cases that should return no restricted material. Keep the decision attached to evidence you can inspect and repeat. A useful plan should explain why moving makes sense, what remains unresolved, and which observations would justify staying with the original. Use that standard to decide whether the managed path earns production traffic.



