Skip to content
NewsProduct launch

AWS ships self-training document agents for Bedrock AgentCore, cutting repeat runs to 90 seconds

· by Pondero Newsdesk

The short version

AWS published a guide showing Bedrock AgentCore agents that write their own document-extraction instructions for new formats and reuse them on repeat runs, cutting processing time from 3 minutes 49 seconds to 1 minute 30 seconds.

AWS ships self-training document agents for Bedrock AgentCore, cutting repeat runs to 90 seconds

A document type that AWS's new Bedrock AgentCore pattern has never seen before takes about 3 minutes 49 seconds to process the first time, and about 1 minute 30 seconds on every similar document after that, because the agents write their own extraction instructions and save them for reuse, per AWS's October 14 walkthrough.

What AWS built

The guide, published on AWS's Machine Learning blog, lays out a multi-agent system on Amazon Bedrock AgentCore that classifies, extracts, validates, and learns from business documents such as invoices, purchase orders, insurance claims, and loan applications, according to AWS. When a document arrives, a Matcher agent checks a vector store for a similar format. A new format routes to a training workflow: an Instructions agent drafts extraction prompts, an Extractor agent pulls structured data, and a Validator agent checks the result against schemas and business rules. If validation fails, an Instructions Fixer agent diagnoses the mismatch, rewrites the prompt, and retries; once the data validates, a Save Instructions agent persists the working prompt to the vector store. The next document of that same type skips straight to the automation workflow and reuses the saved instructions. AWS reports the deployment takes 15 to 20 minutes and the whole workload, 10 training documents plus 1,000 processed documents a month, runs at roughly $58 a month, with foundation-model inference making up about 85 percent of that cost.

Teams skip the template rewrite that breaks on every new vendor layout

Rule-based document pipelines typically need a hand-built template and extraction rules for every document format, and a new vendor's invoice layout showing up unannounced is enough to break the pipeline until someone edits the template. AWS's pattern replaces that manual step with an agent that writes its own extraction rule, tests it against validation, and fixes its own mistake if the first attempt misreads the format, as it did in AWS's own walkthrough example, where a vendor's SKU format did not match the buyer's internal naming until the Instructions Fixer agent added a prefix rule and reprocessed. That shifts the operational burden from "someone maintains templates" to "someone maintains validation rules," which is a smaller and more stable surface. The published cost model also gives finance teams a starting number to scale from: since inference drives most of the spend, cost should move roughly with document volume rather than with the number of distinct formats a team's documents arrive in.

This ships as a reference architecture, a GitHub sample stack a team deploys and maintains itself, not a one-click managed AWS service. A team already running document processing outside Bedrock, or one that wants this behavior without standing up AgentCore, Aurora DSQL, and the included CloudFront dashboard, does not get it for free.

What to watch next

Watch whether this instruction-refinement pattern gets folded into AgentCore's managed tooling directly, rather than staying a reference architecture readers have to assemble themselves. AWS's own framing, that the system "trains itself" on new formats, is also the kind of capability a packaged Textract or AgentCore feature would plausibly absorb within a few release cycles.

Sources