Anthropic Dynamic Workflows: Code-Driven Agent Control

A typical pilot looks like this. The team writes a long prompt for the agent and gives it 40 steps.

Around the thirtieth step, the agent goes off course, and no one can say where the failure began.

The second option is called a “swarm”: a dozen agents freely exchange messages with one another.

The token bill becomes unpredictable, and the entire result must be checked manually. In both cases, the process lacks a framework.

The CFO sees the expenses but not the metric they improved.

After a quarter, management usually shuts down such a pilot and labels its creator an AI-theater enthusiast.

TTU, time to use—the time it takes for a tool to start producing results.

A deterministic skeleton reduces TTU because the result is repeatable.

The same workflow can be measured, compared with manual work, and presented to management as a figure.

Serious engineering lies behind the outwardly simple flow.

To write it, the team breaks the process into steps, defines the input and output for each step, moves validation to a separate stage, and decides what counts as a defect.

Most companies get stuck on exactly this work.

Against this backdrop, choosing a model takes minutes.

In KT.Team projects with Pimcore and Akeneo catalogs, product enrichment is designed as a pipeline from the outset.

The first agent organizes items by taxonomy, the second writes product card copy, and the third builds filter facets; manual tag overrides from category managers are then applied.

A separate checking agent compares the card text with the attributes in PIM.

If the text says cotton but the attribute says polyester, the checker sends the card back for revision, and it never reaches the website. In integrations with 1C and ERP, agents receive data through MCP connectors and have no direct database access.

All traffic to models passes through the LLM & Security Gateway: the gateway determines which fields leave the perimeter and logs every call. In a CIS environment, the same workflow works with GigaChat, YandexGPT, or Qwen: the orchestrator is not tied to a specific model.

Our News Machine, which prepares content for this blog, follows the same structure: source collection, a signal pool, text synthesis, and quality gates.

If the material lacks a strong signal, the gate rejects it and it is not published.

The process is defined as a sequence of steps, each with an input and output schema.

A separate agent performs the check against a measurable criterion.

Limits are set for the number of agents and the cost of each run.

After a failure, the run continues from the last successful step.

It is known which data is sent to the model and where this is documented. If two of the five questions have no answer, flows will only accelerate token consumption.