How Amazon Bedrock AgentCore helps keep technical documentation up to date

Technical documentation often struggles to keep pace with software development. Source code changes continuously as new features, integrations, and dependencies are introduced, while documentation may be updated much less frequently. As a result, developers and other teams can end up relying on information that no longer reflects the actual state of their systems.

AWS demonstrates how an agentic approach powered by Amazon Bedrock AgentCore can help automate the synchronization of technical documentation with source code.

The documentation drift challenge

Documentation drift occurs when technical documentation gradually becomes inconsistent with the actual implementation of a software system.

According to data cited by AWS, documentation can be updated 5.7–25 times less frequently than source code. This creates additional workload for developers, makes systems more difficult to maintain, and increases the time required to understand existing solutions.

The challenge becomes particularly significant in large environments consisting of dozens or hundreds of repositories, microservices, and interconnected components.

An agentic approach to documentation

AWS Advanced Partner Corley developed a solution for Eutelsat that uses an AI agent to automatically analyze source code and help keep project documentation synchronized with it.

A domain expert first populates a knowledge base with relevant documentation and domain-specific information. The user then specifies the repositories that need to be analyzed.

The AI agent accesses the source code, retrieves relevant context from the knowledge base, and compares existing documentation with the actual state of the software. If documentation is missing or contradicts the code, the agent can create or update it.

In this approach, source code acts as the primary source of truth for the current state of the software.

AWS solution architecture

The solution combines several AWS services and technologies.

Domain-specific documentation in Markdown format is uploaded to Amazon S3. Based on a predefined schedule, AWS Lambda synchronizes these materials with Amazon S3 Vectors, making the information available through Amazon Bedrock Knowledge Bases.

When a user initiates documentation synchronization, Amazon Bedrock AgentCore Runtime launches an AI agent built using Strands Agents.

To interact with code repositories, the agent uses Model Context Protocol (MCP). A GitLab MCP server allows the agent to access and analyze source code, while Amazon Bedrock AgentCore Gateway provides connectivity between the agent and MCP servers.

After analyzing the code, the agent retrieves relevant information from the Amazon Bedrock knowledge base to understand domain terminology, system relationships, and dependencies.

Automatically updating documentation

The agent then accesses existing project documentation hosted in Atlassian Confluence through MCP.

It compares three primary sources of information: source code, existing project documentation, and domain-specific context from the knowledge base.

Using this information, the agent can create new documentation or update existing materials.

Instead of periodically rewriting documentation manually, organizations can establish a workflow where AI continuously helps align technical documentation with the actual state of their software.

Results

Eutelsat and Corley tested the solution across 30 repositories that previously had no documentation.

Agent execution, including source code analysis, documentation generation and publishing, took approximately 15–20 minutes per repository. Human review of documentation generated from scratch required approximately 1–2 hours per repository.

All 30 repositories were processed in roughly one calendar week, while AWS costs amounted to only a few dollars per repository.

According to the results presented by AWS, the solution delivered more than 90% estimated time savings compared with a fully manual approach.

Keeping humans in the loop

Automation does not eliminate the need for expert review.

The architecture follows a human-in-the-loop approach. AI performs much of the repetitive work by analyzing code, collecting relevant context and generating documentation, while engineering teams review and validate the resulting materials.

This allows organizations to combine the speed of generative AI with the knowledge and experience of their technical specialists.

Conclusion

Outdated technical documentation is a challenge faced by software development organizations across industries.

Combining Amazon Bedrock AgentCore, Retrieval Augmented Generation (RAG), Amazon S3, AWS Lambda and MCP-based integrations provides a way to automate documentation workflows and keep technical information aligned with the actual state of source code.

The Eutelsat example demonstrates how agentic AI can be applied beyond content generation to practical software engineering tasks, including creating, updating and maintaining technical documentation at scale.

Source: AWS Public Sector Blog — Reducing documentation drift with Amazon Bedrock AgentCore