Agent Dev API

Understand the boundary

An agent workflow can propose or choose actions across several steps. The application still needs to decide which tools exist, which arguments are valid, and which actions require a person’s approval. Model an operation around the task the user requested, not around a generic ability to execute arbitrary commands. Separate read access from the ability to create, modify, or delete.

A practical starting project

Design an agent that drafts an issue from a selected error report. It may read the report, inspect approved project metadata, and prepare a proposed title and body. Creating the issue should be a distinct step with the target repository and final payload shown to the user. Store a durable operation identifier so a retry does not create a second issue.

Where integrations go wrong

Do not delegate authorization to a prompt or accept a model-supplied tenant identifier as authoritative. The tool handler should derive identity from the authenticated application context. Define a maximum amount of work, a timeout, and an explicit incomplete state. A controlled stop is better than an agent repeatedly guessing its way through a failure.

Reference for implementation: OpenAI function calling documentation. Check the official reference for the specific version and environment you plan to use. The exercise above is an engineering starting point, not a live DevAPI.com service.