Field Guide

Working With Agents

A field guide for professionals who direct AI — not just use it. Directing agents well is a skill. Not prompt engineering. Something deeper.

By Laura Tomas Jimenez·Managing Disruptions·25 min read

Part 01

What Agents Actually Are

This is not a prompt engineering tutorial. It is a guide for professionals who have realised that using AI is not the same as working with it — and who want to get serious about the difference.

An AI agent is a system that can take a goal, break it into steps, execute those steps, evaluate the results, and adjust. Unlike a simple chatbot, an agent interprets your goal and makes decisions about how to get there.

This is why directing agents is a skill, not just a feature you turn on.

Agents are strong at

Structured tasks with clear success criteria
Research and synthesis across large volumes
First drafts of standardised documents
Data analysis and anomaly detection
Rapid iteration — ten variations at once

Agents struggle with

Knowing when the brief is wrong
Applying context you have not articulated
Recognising emotional dynamics
Decisions with ambiguous long-term consequences
Genuinely novel ideas outside their training

Part 02

The Three Layers of Directing Agents

Most people interact with agents at one layer. Professionals who get exceptional results operate at all three.

Layer 01

The Brief

What most people think of as prompting. But a brief is more than a prompt. A good brief includes context the agent cannot infer, quality criteria — not just format specs — and examples of what good looks like. And what bad looks like.

The key distinction

Format spec: "Write a professional email." Quality criterion: "Write an email that acknowledges the client's frustration without accepting liability, maintains warmth, and opens a path to resolution." The gap between these two is where mediocre agent output comes from.

Layer 02

The Evaluation

Where most people fail. They receive output and accept or reject it — binary. Professional direction requires structured evaluation across three passes: factual accuracy, judgment quality, and what is missing.

The most important pass

The third pass — what is absent — is the one most people skip. What did the agent not say that it should have? What context did it lack? What questions does the output raise that the agent did not anticipate?

Layer 03

The Iteration

This is the layer that turns passable output into excellent work. Treat the interaction as a conversation, not a transaction. Give specific feedback, not general dissatisfaction. Translate your intuition into observations the agent can act on.

Translate your gut

"This doesn't feel right" is useful for you, not for the agent. Convert it: "The tone in paragraph three is too formal for this audience" or "You emphasised cost savings but the real value proposition is risk reduction."

Part 03

Building Agent Workflows

Pick one workflow from your week and redesign it around agent collaboration. For each step, classify it:

Mechanical — Clear inputs, clear outputs. Agent handles this.

Evaluative — Comparing output against standards. You review, agent produces.

Judgment — Requires context or experience. You handle this.

Relational — Requires human presence or accountability. Agent cannot touch this.

Part 04

Common Mistakes

Treating agents like search engines

Asking a question and accepting the first answer. Agents are generative systems, not retrieval systems. Their first output is a draft, not a fact.

Fix: Treat every output as a starting point for evaluation, not an endpoint.

Under-briefing

Giving a vague instruction and being disappointed by a vague result. The agent is not reading your mind. It is working with what you gave it.

Fix: Invest time in the brief. Include context, criteria, examples, anti-examples. The brief is the work.

Over-trusting after good outputs

After a few good results, assuming the agent "gets it." Consistency is not reliability. Agents can produce excellent work ten times and confidently wrong work the eleventh.

Fix: Never skip evaluation. Build a checklist specific to your domain. Use it every time.

Doing the agent's job better

Spending more time fixing output than it would take to produce it yourself. If you are consistently rewriting more than 30%, the problem is upstream.

Fix: Go back to the brief or the task assignment. The issue is never in the editing.

Ignoring the identity question

Using agents effectively without addressing what it means for how you see your professional self. The tools only matter if you know who you are using them.

Fix: Read The Human Edge guide. This is the foundation everything else sits on.

Part 05

Your First Week Protocol

One workflow. Five days. A reflection worth more than any certification.

Day1

Choose one recurring task

Pick something you do at least weekly. Clear inputs and outputs. Not the most important thing you do — something mid-stakes.

Day2

Write the brief

Before touching any agent: What is this task? Who is it for? What does good look like? What are the three most common ways this goes wrong?

Day3

Run it

Give the brief to your agent. Do not intervene during production. Let it produce a complete output.

Day4

Evaluate ruthlessly

Three-layer evaluation: factual accuracy, judgment quality, what is missing. Document everything you would change and why.

Day5

Iterate

Feed your evaluation back. Run it again. Compare the second output to the first. What improved? What did the agent still miss?

Endof week

Reflect — on yourself, not the tool

What did this reveal about your expertise? What do you know that you could not articulate before? Where is your judgment most valuable — and where were you just performing tasks?

Ready to go deeper?

Managing Disruptions offers AI Strategy Sprints and Team AI Training built around these principles — not tool tutorials, but the judgment and workflow design that make AI actually work.

Work with Laura Read The Human Edge