Why agentctl

The useful boundary is not “agents instead of automation.” It is deterministic automation that can delegate a bounded judgment task without giving the model control of authority, graph structure, persistence, or recovery.

The problem with ordinary agent scripts

A direct model call is easy to start. It becomes harder to review when prompts imply authority, tool schemas drift, retry can duplicate an effect, an interrupted process loses progress, or a pipeline needs one stable result. Adding ad hoc state and approval logic can turn a small script into a hidden workflow engine.

agentctl makes those concerns explicit in one strict document and one durable runtime:

• declaration-ordered graph compilation;
• typed inputs, outputs, templates, actions, agents, and tools;
• policy checks outside model prompts;
• effect identity and conservative uncertainty;
• durable approvals, checkpoints, audit events, and trace correlation;
• resume, recorded replay, and explicit fork semantics;
• native provider adapters behind one neutral core.

Related approaches

Approach	Choose it when	Add agentctl when
Shell script calling a model	The task is small, disposable, and has no recovery or policy need.	The call becomes one step in a reviewed graph with effects and durable evidence.
Chat-agent framework	Interaction and dynamic conversation are the product.	A reviewed workflow, not a conversation, must own control flow.
General workflow engine	You need broad scheduling, distributed workers, or many integration types.	A task needs bounded model or tool semantics and a local correctness record.
Ansible	You need configuration management and convergent idempotent resources.	You need a bounded reasoning step, not a replacement for configuration management.
CI/CD pipeline	You need runners, triggers, secret injection, and artifact retention.	One pipeline step needs strict agent workflow execution and recovery semantics.

Reasons not to choose it

Do not choose the current release when you require a 1.0 CLI or long-term support contract, distributed execution, a hosted scheduler, a visual builder, an operating-system-enforced in-process sandbox, or live support evidence for every provider. Read Limitations before adopting workflow API agentctl.dev/v1.