#team

An Agent in Production: Boring Commits Matter More Than Demos

A review of Microsoft Agent Framework dotnet-1.24.0 and choosing between Document Intelligence and Content Understanding: where AI projects lose results.

  • A Release Without Headline Features
  • What Is Included in the dotnet-1.24.0 Release
  • Why Null in a Session Deserves Its Own Commit
  • Security Lives in Validation

A Release Without Headline Features

The dotnet-1.24.0 release of Microsoft Agent Framework contains almost no headline features: preserving null in session state, validating identifiers, and checking capabilities for code execution. During the same period, the Foundry team published an analysis of how to choose between Azure Document Intelligence and Azure Content Understanding. Both pieces show what engineers assemble an AI implementation from: details absent from demos that determine whether a project will make it to the second quarter.

What Is Included in the dotnet-1.24.0 Release

The changelog reads like a list of defects found only under the load of real-world scenarios.

Some of the items include: preserving explicit null in session state during round trips (#9012); accounting for disabled recent-search memory (#9013); preserving lazily created request messages when calling the agent (#9017); rejecting sensitive identifiers in declarative descriptions (#9071); strengthening capability checks in LocalCodeAct (#9057); stabilizing the package for Anthropic agents (#9111); and restoring the decision filter for a source-only release (#9140). Individually, these are minor details.

Together, they distinguish a prototype from a system that can be put into production.

Why Null in a Session Deserves Its Own Commit

The agent retains conversation state between calls. If, after serialization, “the user explicitly chose nothing” and “no value was provided” become indistinguishable, the agent starts behaving differently than it did an hour ago. The user says, “AI is acting strange today.” The engineer gets an irreproducible bug because the state looks normal in the logs. The same applies to the disabled-memory flag. An administrator turned off recent-query storage, but the agent continues using those queries: that violates compliance requirements.

If customer data is stored in 1C or EDI, the security team will not approve such a solution.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

Security Lives in Validation

Two release items concern protection: rejecting sensitive identifiers in declarative configurations and checking capabilities in LocalCodeAct, where the agent executes code.

The rule is simple: the more actions an agent performs on its own, the more strictly its permissions must be checked.

An agent configuration in YAML or JSON is also an input interface and must be validated like user input.

That Is Why We Build a Control Layer Around Agents

LLM & Security Gateway sits between the model and corporate systems, filters data, and records exactly what the agent requested.

The framework addresses some risks internally, while the perimeter remains the integrator’s responsibility.

Documents: Tool Choice Determines the Cost of Ownership

  1. A Microsoft Foundry article notes that document processing looks simple while all input files have predictable formats. In reality, invoices arrive in different layouts, material contract terms are hidden in prose, and a single business process may combine PDFs, Office documents, images, audio, and video.

  2. The authors identify the criteria for choosing between Azure Document Intelligence and Azure Content Understanding: extraction quality; labeling effort; latency and cost; deployment options; reasoning and source grounding; and the amount of pipeline code that will need to be maintained.

  3. Both services rely on shared core content-extraction capabilities.

  4. The difference is how much work remains for your team: labeling samples, adding wrapper code, and explaining the result to an auditor. ###

  5. How to Apply This to a Purchasing Decision

  6. We recommend measuring time to first useful result (TTU, time to use) and support costs after one year.

  7. A pipeline with hundreds of lines of wrapper code around a recognition service is cheap in the first month and expensive by the third, when a major provider changes its invoice format.

  8. This is our assessment based on integration projects, not a conclusion from the Microsoft article: the text contains no figures on cost of ownership.

Where the Two Stories Converge

  1. The agent framework release and the analysis of document services raise the same central question: how much engineering work lies behind a simple result.

  2. To make the interface look simple, developers must preserve state, check permissions, normalize formats, and display the answer’s source.

  3. That is why we start AI-native integration with contracts, trust boundaries, and measurable outcomes, and choose the model afterward. An agent that reads documents through RAG or accesses systems through MCP servers inherits weaknesses from both layers: it loses state, receives excessive permissions, or extracts a contract field incorrectly.

  4. An error at any stage affects the business metric.

Verdict

When an executive asks whether an AI project is ready to launch, open the changelog of a mature framework and count how much wrapper code will need to be maintained. In our experience, a release containing fixes for state, memory, and validation says more about maturity than an announcement of a new model. AI that does not move a metric remains theater, paid for by the people who approved it.

Discuss the article: Agents in Production: Boring Commits Matter More Than Demos

Enter your email or phone number so we can get back to you.

Send via:

By clicking the button, you accept the offer and grant consent to the processing of personal data.

Related solutions

All solutions
AI monitoring is automatic control of processes and eventsMonitoring construction, payments, and salesAI assistant for Rosreestr: EGRN extracts, verification, and monitoringDelivery quality control for retail

Articles on the topic

All articles

Related videos

All videos
Shorts Why cause failures in production? DevOps
Shorts Why do test environments hold engineers' growth back? DevOpsEngineering approach
Shorts Why do rare releases break quality? DevOps

News on the topic

All news