← Back to Blog

Plans, Not Prompts: Why Agent Frameworks Need an Execution Model

· Zakardo

The current generation of AI agent frameworks treats agent behavior as an emergent property of prompt engineering and tool calling. An agent is given a set of tools, a system prompt, and a goal. It reasons step by step, calls tools, and produces a result. The developer’s job is to write tools, craft prompts, and hope the agent behaves correctly.

This model has fundamental problems.

No Program Exists

There is no inspectable, versioned, testable artifact that describes what the agent will do. The agent’s behavior is a function of the LLM’s reasoning at inference time. Two identical requests may produce different execution paths. There is no way to review, approve, or audit what the agent will do before it does it.

No Governance

When an agent calls an external API, sends an email, or modifies a database, there is no policy layer that evaluates whether that action should be permitted in context. Side effects are fire-and-forget. The developer must build guardrails manually, per tool, per use case.

The Gap

The gap is not “better tools for agents.” The gap is a missing execution model.

Current frameworks solve: How do I give an LLM the ability to call functions?

The question that needs answering: How do I build, govern, validate, and trust the programs that agents execute?

This is a runtime problem, not a prompting problem.


Zakardo introduces a new execution model where agent behavior is expressed as structured, versioned, validated plans, composed of governed, typed units of execution called capabilities. Plans are first-class programs — the core artifact, not an implementation detail.

Get started with Zakardo →