Agent Framework For Productivity

Introduction

There is a large divide between how developers and the rest of us are using AI.

We are seeing engineering change drastically in 2025. From ‘vibe coding’ to ‘context engineering’, these buzzwords and real practices are improving the productivity of developers 10x (or maybe even 100x)! Full application prototypes can now be written in minutes. High functioning product teams can ship fixes on the spot, rather than days or weeks later.

Yet, I found myself (and many of my peers) still writing long prompts and copying/pasting results between applications.

I want developer-like scale in my life too!

So I set out on this mission.

Why is a framework necessary?

Implementing a framework for your agent workflows has multiple benefits, namely: consistency, reliability and efficiency.

• Consistency for agents removes ambiguity & significantly improve an agent’s ability to match a request to the right task
• Consistency for you to make tweaks and not lose your mind. Everything is organized in a hierarchical and transparent manner
• Efficiency of time & cost :
  • Getting it right it a non-negotiable goal of any AI system should be getting the right result on the first request
  • Context management is about the right amount of information, at the right time - keeping tokens (cost & confusion) optimal (see: Context is Everything )

---

Framework

The framework replicates a lot of the unwritten parts of how an organization/team operates. It starts with a mission and flows down to management and then to granular tasks.

You get out what you put in.

---

Governance

Purpose: Describes mission-level goals and guide agents on how to interpret requests and how to fulfill them. Enough information should be included to successfully ‘onboard’ an agent, without too much detail becoming overwhelming. These instructions should always be read first by every agent.

• Mission: What we must achieve together
• Policies: The rules by which we operate together
• Procedures: Overview of common tasks/workflows and how the agent and user (both) approach these. Each of these are defined in a separate Directive file)
• Context: Background information required for every task. These are the standards that apply to everything and agent (and user) does

Examples:

• your mission, who you are, what role the agent should play (most people are doing this part already today)
• an index of all your directives (links to separate files)
• culture and supporting goals
• folder structure
• file naming requirements
• voice & tone
• universal glossary

Practical Tips

Think like a CEO. Imagine you are building a company with the sole purpose of doing your job today. Who would you hire? Your role is to provide vision and clarity, not to get into the weeds. The better you can orchestrate, the more effective delegation will work. You’ll need to get comfortable with this separation.

These high-level instructions are the critical foundation to understanding how work gets done. Your core governance model shouldn’t change that often, with the exception of adjusting Directives.

---

Directives

Purpose: Describe individual tasks or workflows (combination of tasks) that are available for an agent to execute. These instructions are read on as-needed basis giving extra detail to that specific request, task, and deliverable required. They are outcome-driven and easy to read. A great directive is written so that almost anyone can complete the task. When things don’t go well, it’s probably due to ambiguity or unclear definition at this level.

• Mission: What the task or workflow must do
• When to execute: When an agent should consider (and follow) this directive (i.e. what do typical requests from the user look like?)
• Task details: The detailed steps performed by this task:
  • What should the input(s) look like
  • What additional data is read?
  • How is the information processed? By the LLM? By a local Tool? By a remote server or API?
  • What should the output(s) look like?
• Success criteria: No ‘hoping’! What does success look like at different steps in the process? Specific examples here are critical !
• Context: Additional information the agent needs to know about the task or workflow. ex “Is it part of a broader workflow?”
• Known issues: How the agent should handle exceptions (also see Feedback )

Examples:

• Task directives (min. one directive per task)
• Workflow directives that string together multiple tasks (you can mandate the workflow or provide options/guidances)
• New Directive directive - how an agent will handle a request to create new repeatable process
• New Tool directive - how an agent will handle a request to create a new tool
• Additional Framework Directives discussed under Feedback section

Practical Tips

Think like a Manager. At this level, details matter. Your role is to articulate how work should be executed, what success looks like, and what happens if things go sideways.

The greatest level of ‘tuning’ occurs at this level. Sometimes changes to directives come top-down from you manually. Sometimes, they are bottoms-up, automatically applied by agents through Feedback .

---

Execution

Purpose: Define and code executables which take consistent inputs and produce consistent outcomes. At this level, it’s important to describe what the executable does and doesn’t do, and what data in/out will look like. The Execution phase is critical because it offloads work from the agent. I want to MINIMIZE the work my AI agent does. This results in fewer tokens, which means less agent confusion, faster results, and less cost. If we could build everything as an executable, we probably would!

• Documentation: what the executable does, what inputs must look like, what outputs will be, etc.
• When to execute: When an agent should consider (and call) this executable (i.e. what do typical requests look like)
• Context: Additional information the agent might need to know about the broader task or workflow that could call it

Examples

• Scripts (Python, etc.)
• MCP Server
• API calls

Practical Tips

Think like an Engineer. Your role here (and eventually something you can delegate to an Agent) is to create executables that can take standardized inputs and produce standardized results. The goal here is bite-sized work that can be executed with consistency. If something can be fully automated without an AI agent - why not?

What value is your AI agent providing over traditional automation? <- The answer is the line between Execution and Directives .

---

Feedback

Purpose: Describe controls and procedures for how we handle improvement to the system. No employee, system, dataset, etc. will be perfect on day 1 - and every day going forward. The same mindset should be offered to your agents. This part of the framework focuses on how to handle quality checks, edge cases, changes, and learning. It also enables you to add new capabilities that future agents can use.

• Quality checks: within every Directive - validating outputs against baseline examples you have provided
• Framework Self-Ameliorating Culture: every agent’s last task is to look for ways to improve its instructions for the next agent, essentially: learning and self-improving
• Learning: such as adding new capabilities for the agent

Practical Tips

This is an often over-looked (and difficult) element to implement. Also because organizations struggle to get this right too. When implemented well, this element is a largely (automated) feedback mechanism to modify Directives and Executions.

The Feedback layer constantly re-evaluates & changes Directives when execution doesn’t operate correctly. It can modify the Execution layer to change (or add new) steps to ensure outputs meet expectations.

---

Summary

Implementing a framework, will drastically improve productivity by driving quality outcomes and reducing the amount of information you need to give to get there.

In the next post, I’ll be sharing specific implementation examples and sample scripts.