Agent SDKs in business workflows: adaptive steps, fixed conditions

Question

Can a coding loop become a monthly reporting workflow?

An AI preparing a monthly report finds the documents, cleans up tables, and produces attractive charts. The accounting reviewer still cannot tell whether its revenue figure is comparable with last year’s. Does it subtract returns, or represent the value of orders received? This scenario from the article illustrates the difference between making progress and preserving the conditions of the work.

The September 29 Claude Agent SDK explanation renames the Claude Code SDK and describes reusing its machinery beyond coding. It handles files and commands through a loop of gathering context, acting, and checking work. That loop offers a more useful starting point than names of industries it might serve.

The question here is what the AI should choose and what should remain fixed when that loop enters a business process. My judgment is to allow adaptation in finding information and resolving gaps, while keeping business definitions and calculation conditions in explicitly defined operations. Reuse becomes valuable when the AI can select the necessary steps around business rules without having to invent those rules.

Reusable machinery

Intermediate findings change the next action

The SDK explanation includes file search and editing, commands, and external tools. It also describes subagents with separate context and compaction for longer work. These components reduce the need to rebuild machinery for choosing the next action from intermediate results.

In the reporting scenario, unexpected column names might lead to a search for explanatory documents; missing values might lead to another data source. A loop has a role when every branch cannot be written in advance. Yet freedom to change a search path differs from freedom to change the definition of revenue. Allowing the first does not require delegating the second.

Iteration benefits from concrete checks. The SDK explanation discusses rules, visual feedback, and model-based judgment, while qualifying the robustness of the latter. A chart that looks plausible does not establish which transactions belong in its calculation. Applying a loop to business work requires specifying which check can detect each relevant failure.

Tool design

Define the unit of work before adding more APIs

The quality of checking also depends on what tools return. Anthropic’s tool-design explanation recommends searching contacts rather than listing all of them, and consolidating relevant customer information rather than retrieving customers, transactions, and notes separately. The approach avoids filling limited context with irrelevant intermediate results and aligns available actions with the task.

For monthly reporting, a tool that freely queries every database table presents different choices from one that retrieves an approved metric for a specified period and region. The latter can still receive a misinterpreted request. However, keeping a shared organizational calculation inside the tool reduces the scope for reconstructing it on every run.

A consolidated tool can also conceal its internal work. Returning only an amount while hiding a failure to retrieve returns data makes a short response misleading. The metric used, retrieval coverage, and missing inputs need to remain available because they affect the next step. This is the article’s design recommendation; it is not a claim that shorter responses always improve accuracy.

Fixed meaning

Separate the revenue definition from generation on each run

The Looker MCP Server announcement describes connecting AI to Looker’s semantic layer, where metrics have shared definitions. Looker generates SQL from those definitions rather than leaving the AI to improvise it. The execution loop can interpret the request and select a tool while an existing definition supplies the calculation.

Selecting a defined metric still does not finish the work. Does domestic revenue refer to the destination country or the location of the selling company? Does the period use order dates, invoice dates, or recognition dates? In this scenario, candidate metrics must be matched to the purpose of the request, with clarification where ambiguity remains.

Even a correct metric can produce a misleading comparison if last year’s figure used another definition. Recording the metric, period, regional conditions, and definition version consulted would help. Retrieving and recording the version is an application design proposed here, not a guarantee that Looker MCP returns all these fields together. This arrangement fixes the calculation’s meaning; it does not establish what caused revenue to change.

Limits of checking

Valid syntax does not establish the intended behavior

Tools that update data make this distinction concrete. Google’s Firestore MCP Toolbox explanation covers queries, targeted document updates, and security-rule validation. Its story about Alex investigating 20 test accounts and removing obsolete wishlist items is expressly hypothetical. The concluding reduction from hours to minutes cannot be treated as a measured deployment result.

In that example, rule validation identifies a missing semicolon. Alex then retrieves the active rules to compare them with her new rules. Syntax repair and rule comparison are separate steps. Passing syntax validation does not establish that only intended users can read documents or that only the correct product was removed.

Reporting likewise needs separate checks for executable SQL, conformity to the calculation definition, and suitability for the request. Passing one cannot serve as evidence for passing the others. Tools can then restrict permitted inputs and update scope, returning errors when conditions cannot be met. Retrying should supply missing information to satisfy the same conditions, rather than search for a way around them.

Dividing execution

Use adaptive exploration inside a defined sequence

This division also helps compare SDKs. The October 1 Microsoft Agent Framework announcement distinguishes model-driven Agent Orchestration from business-logic-driven Workflow Orchestration. It describes a direction in which open-ended work can coexist with defined procedures.

In the monthly-report scenario, fix the sequence: confirm the period and metric, aggregate under that definition, check the calculation, prepare a draft, and submit it to a reviewer. Leave room for the AI to choose how to find information or write an explanation within each stage. If returns data is missing, stop with the gap rather than treating it as zero. If another search cannot resolve the gap, hand it to the responsible person instead of completing the report with the same incomplete amount.

The reviewer would see more than the final prose: the metric and coverage, calculation checks, and unresolved questions explain whether submission conditions hold. Whether this reduces human effort still needs testing. A large record alone does not ensure fewer missed problems. Present what has been checked automatically and what requires human judgment alongside the evidence for the report.

Microsoft also describes integration with the Microsoft 365 Agents SDK and a shared cloud runtime as future directions. Those statements cannot all be read as current default capabilities. Other SDKs need not offer the same division of execution or stopping behavior; check the version actually being used.

When to decline

A known calculation does not need more iteration

This does not justify adding an SDK to every task. Anthropic’s agent-design explanation also recommends starting simply and adding complexity when needed. If the job only aggregates identically formatted files under the same definition each month, an existing script with exceptions handed to a person remains an option.

An SDK has a stronger case when information formats or locations change and intermediate findings must guide the next action. That differs from asking a model to reconstruct the same calculation repeatedly. Compare more than processing time: examine wrong metric selections, the proportion of missing-data cases that stop correctly, and the time staff spend investigating again. Keep the option to return work to a fixed procedure if the added freedom does not improve outcomes.

Judgment

Generalize how work proceeds while preserving its conditions

Claude Agent SDK, business-data connections, and Microsoft Agent Framework expand reuse from different parts of execution. Their features and use cases do not establish that an entire business process can be delegated. The comparison instead suggests more room to design exploration, definitions, and fixed procedures separately, then combine them.

For monthly reporting, give the AI the task of finding gaps and seeking clarification without letting it rewrite revenue definitions or submission conditions inside its loop. That division gives a coding mechanism a meaningful role in another job. Judge the SDK by how concretely the team can define what may be delegated, rather than by the number of possible uses.